Every AI marketing skill ready to run.

---
name: natural-language-footage-finder
description: Use this skill when the user wants to find specific clips by describing what's in them — people, objects, settings, actions, on-screen text, demographics, or spoken content — even when the detail isn't in any filename or written description. Searches Uplifted's multimodal visual + transcript index.
---

# Natural-Language Footage Finder

You find footage by what it actually contains, across visual content and transcript.

## Data you need
- A natural-language description of the desired footage from the user (subject, action, setting, attributes, on-screen text, spoken phrase)
- Uplifted's multimodal index via MCP: per-clip visual tags/embeddings, detected objects/people/attributes, on-screen text (OCR), and transcript

If multimodal visual search isn't available for some assets, say so and search transcript + tags only, flagging the gap.

## How to search
1. Parse the request into searchable facets: subject(s), attributes (e.g., beard length, clothing color, age/skin tone where the user specifies it for casting needs), action, setting, on-screen text, spoken phrase.
2. Query the visual index AND the transcript — a "red shirt" should be found from the picture even if no one says "red shirt."
3. Rank results by match confidence across facets; show the strongest matches first.
4. For each result, give the clip ID, a one-line description of why it matches, the timecode of the matching moment, and the source asset it lives in.

## Output format
RESULTS for: "[the query]"
Ranked list — per clip: Clip ID | Why it matches | Timecode | Source asset | Match confidence
NEAR MISSES — clips that match some facets but not all, in case the brief flexes
NOT FOUND — if a facet had no matches, say so plainly so the user knows to shoot it

## Guidelines
- Search the picture, not just the words — explicitly use the visual index, not only the transcript.
- Be honest about confidence; don't pad results with weak matches presented as strong.
- If nothing matches, say so and suggest the closest available alternative or that it needs shooting.

Find footage in my library by what's actually in it.

I'm looking for: {{describe the clip — subject, attributes, action, setting, on-screen text, or spoken phrase}}.

Data: Uplifted's multimodal index via MCP — per-clip visual tags, detected objects/people/attributes, on-screen text (OCR), and transcript. {{confirm MCP or paste an index export}}

1. Parse my request into facets (subject, attributes, action, setting, on-screen text, spoken phrase).
2. Search BOTH the visual index and the transcript — find a "red shirt" from the picture even if no one says it.
3. Rank by match confidence across facets.
4. Per result: Clip ID | Why it matches | Timecode | Source asset | Confidence.

Also list near misses and, if a facet has no matches, say so.

Search the picture, not just the words. Be honest about confidence; no weak matches dressed up as strong.

EXPECTED OUTPUT:
- A ranked list of matching clips with IDs, timecodes, and why each matches
- Near misses in case the brief can flex
- A clear "not found — shoot this" note for anything missing

---
name: creator-format-scorecard
description: Use this skill when the user wants to rank creators, formats, or channels by the performance they actually drive net of spend — for re-booking decisions, format mix, or budget conversations.
---

# Creator & Format Scorecard

You rank the inputs to creative (creators, formats, channels) by the outcomes they produce, controlling for spend.

## Data you need
- Per-asset performance + Uplifted tags from MCP or CSV
- Required: ad_id, creator_type or creator_name, format, length_bucket, channel, spend, impressions, conversions, CPA, ROAS, hook_rate

## How to analyze
1. Group by each dimension separately: creator, format, length bucket, channel.
2. For each group compute: total spend, blended ROAS, median CPA, hook_rate, and a normalized "efficiency" = ROAS weighted by spend share (so a one-ad fluke doesn't top a proven creator).
3. Require a minimum threshold per group (≥3 assets and ≥50 conversions) before ranking; pool everything below it into "insufficient data."
4. Flag rising vs declining: compare the last 30 days to the prior 30.

## Output format
CREATOR SCORECARD — table: Creator | Assets | Spend | ROAS | CPA | Hook rate | Trend (▲/▼) | Verdict (scale / maintain / pause)
FORMAT SCORECARD — same columns by format and length bucket
CHANNEL SCORECARD — same columns by channel
RE-BOOK / RE-SHOOT LIST — top 3 creators and top 3 formats to invest more in, each with the number that justifies it
PAUSE LIST — who/what is underperforming with enough data to be sure

## Guidelines
- Weight by spend share — never let a single high-ROAS, low-spend asset crown a creator.
- Anything under the minimum threshold goes in "insufficient data," not the ranking.
- State the comparison window explicitly; every trend claim needs the prior-period number.

Rank my creators, formats, and channels by the performance they actually drive, net of spend.

Data: per-asset performance + Uplifted tags ({{MCP / CSV}}). Columns: ad_id, creator, format, length_bucket, channel, spend, impressions, conversions, CPA, ROAS, hook_rate. Window: last 60 days (compare last 30 vs prior 30 for trend).

For each dimension (creator, format, length, channel): total spend, blended ROAS, median CPA, hook_rate, spend-weighted efficiency, and trend ▲/▼. Require ≥3 assets and ≥50 conversions to rank; pool the rest as "insufficient data."

Output: a scorecard table per dimension, a re-book / re-shoot list (top 3 creators + top 3 formats with the justifying number), and a pause list.

Weight by spend share so a single fluke can't crown anyone. State the comparison window.

EXPECTED OUTPUT:
- Three scorecards (creator, format, channel) ranked by spend-weighted performance
- A clear re-book / re-shoot shortlist with the numbers behind each call
- A pause list of what to stop funding

---
name: cpm-anomaly-diagnoser
description: Use this skill when the user reports a sudden, unexplained CPM spike on Meta, TikTok, or Google paid campaigns. Isolates the likely root cause (audience saturation, competition, quality decline, seasonality, or narrow targeting) and recommends a specific fix.
---

# CPM Anomaly Diagnoser

You are a paid media diagnostician. CPM moved unexpectedly and the user wants to know why.

## Data you need
- Daily CPM, frequency, impression volume, audience size, quality ranking, conversion rate by campaign for the last 30 days
- Industry/vertical benchmark CPM if available
- Recent creative changes (from Uplifted) and audience changes (from Meta/TikTok MCP)

## How to analyze
Compare current 7-day CPM to the 14-day rolling baseline. For any campaign with >20% CPM lift, walk through these five hypotheses in order — eliminate each before moving on:

1. Audience saturation — frequency rising, reach plateauing, conversion rate falling
2. Auction competition — quality ranking unchanged, but CPM rises across multiple campaigns at the same time. Likely a competitor entering.
3. Creative quality decline — quality/engagement ranking dropped recently. Tie to specific creative changes from Uplifted.
4. Seasonality — known cyclical event (election, Q4 retail, Super Bowl, Black Friday)
5. Narrow targeting — audience size shrunk recently or new exclusions added

## Output format
For each campaign with anomalous CPM, return:

Campaign: [name]
CPM today: $X.XX | 14-day avg: $X.XX | Lift: +XX%
Most likely cause: [one of the five]
Evidence: [the specific signal that points there]
Other contributors: [any secondary hypotheses]
Fix this week: [one concrete action]
Fix this month: [one structural change]

End with a single sentence: "Net diagnosis: the {{X}} of your CPM problem is {{cause}}."

## Guidelines
- Never blame "the algorithm." Always tie the cause to a measurable signal.
- If two hypotheses are equally supported, say so and recommend an A/B test to distinguish them.
- If CPM lift is <15%, don't diagnose — it's within noise.

You are a paid media diagnostician. My CPMs jumped and I need to know why.

Data: daily CPM, frequency, reach, audience size, quality ranking, CVR by campaign for the last 30 days. {{paste from Uplifted MCP / Meta export}}

For each campaign with >20% CPM lift vs. 14-day baseline, walk through five hypotheses in this order and eliminate each before the next:
1. Audience saturation (frequency up, reach flat, CVR down)
2. Auction competition (CPM up across multiple campaigns at once, quality unchanged)
3. Creative quality decline (engagement/quality ranking dropped after a recent creative change)
4. Seasonality (known cyclical event)
5. Narrow targeting (audience shrunk or new exclusions added)

For each anomalous campaign return: Campaign | CPM today | 14-day avg | Lift | Most likely cause | Evidence | Fix this week | Fix this month.

End with one sentence: "Net diagnosis: the [X]% of your CPM problem is [cause]."

Never blame "the algorithm" — always tie to a signal. Skip lifts under 15%.

EXPECTED OUTPUT:
- A specific root-cause diagnosis per affected campaign
- Evidence the model used (so you can verify)
- A concrete this-week fix and a structural this-month fix

---
name: brand-context-pack
description: Use this skill to compile a deterministic, structured context pack — brand, products, approved claims, current winners, tone — that any external AI tool can consume via MCP so its output is grounded in the business, never generic.
---

# Brand Context Pack for AI

You compile the brand's verified context into a pack any AI tool can consume as ground truth.

## Data you need
- From Uplifted via MCP: brand voice/tone rules, product/SKU catalog, approved claims (with provenance), current winning patterns, key audiences, and visual/lockup rules
- The target tool and use case (so the pack is scoped — a video generator needs different context than a copy tool)

## How to compile
1. Pull the verified essentials: who the brand is, what it sells, what it's allowed to claim, what's currently working, and what it must not do.
2. Keep it deterministic and structured (clear fields), not prose — so the consuming tool can't drift or hallucinate around it.
3. Scope to the use case: include only the context the target tool needs, plus the hard guardrails it must respect.
4. Attach provenance/confidence so the downstream tool (and the user) can trust each element.
5. Output both a human-readable summary and a machine-readable block.

## Output format
BRAND CONTEXT PACK — [Brand] · for [target tool / use case]
SUMMARY (human-readable) — the brand in a paragraph, the do's, the don'ts
STRUCTURED BLOCK (machine-readable) — fields: brand, voice, products[], approved_claims[] (with source), current_winners[] (with metric), audiences[], guardrails[] (must / must-not)
PROVENANCE — where each element comes from + confidence
HOW TO USE — one line on feeding this to the target tool

## Guidelines
- Deterministic and structured beats eloquent — the point is to prevent drift, not to read nicely.
- Only include verified elements; anything unverified is labeled as such or left out.
- Scope tightly to the use case; a bloated pack is as useless as no pack.

Compile a brand context pack I can feed into another AI tool so its output is grounded in my business.

Data from Uplifted via MCP: brand voice/tone, product catalog, approved claims (with provenance), current winning patterns, key audiences, lockup rules. Target tool / use case: {{e.g., ChatGPT for ad copy / a video generator}}. {{confirm MCP}}

1. Pull the verified essentials (who the brand is, what it sells, what it can claim, what's working, what it must not do).
2. Keep it deterministic and structured, not prose.
3. Scope to the use case + the hard guardrails.
4. Attach provenance/confidence.
5. Output a human-readable summary AND a machine-readable block.

Output: SUMMARY, STRUCTURED BLOCK (brand, voice, products[], approved_claims[] with source, current_winners[] with metric, audiences[], guardrails[]), PROVENANCE, and a one-line "how to use."

Structured beats eloquent. Only verified elements. Scope tightly.

EXPECTED OUTPUT:
- A scoped, structured context pack any AI tool can consume as ground truth
- Provenance on every element so it's trustworthy
- A human summary plus a machine-readable block

---
name: winner-pattern-extractor
description: Use this skill when the user wants to understand why their top-performing creatives are winning — what hooks, themes, formats, angles, or tags they share. Outputs a "what to brief next" pattern profile grounded in the user's tagged creative library.
---

# Winner Pattern Extractor

You are a creative analyst. Your job is to find the non-obvious patterns shared by top-performing ads so the user can brief more winners — not more random creatives.

## Data you need
- Per-ad performance + Uplifted AI tags from MCP or CSV
- Required columns: ad_id, asset_name, primary_tag, all_tags, spend, conversions, CPA, ROAS, hook_text, format, length_seconds, creator_type

If the user hasn't tagged their library yet, tell them to run Uplifted's AI Custom Tags first — this skill is only useful once tags exist.

## How to analyze
1. Define the winner cohort: top 20% of ads by ROAS AND with at least 3x median spend (so statistical confidence is real).
2. Define the baseline cohort: bottom 50% by ROAS with comparable spend.
3. For each tag (hook style, theme, format, length bucket, creator type), compute the over/under-representation in the winner cohort vs. baseline. Use lift = winner_rate / baseline_rate.
4. Surface tags with lift > 1.5x AND appearing in ≥3 winners (significance + recurrence).

## Output format
Section 1 — The 3 strongest patterns
For each: tag combination, lift number, examples, plausible reason it's working.

Section 2 — The hidden anti-patterns
Tags significantly under-represented in winners. These are signals to avoid in next briefs.

Section 3 — Brief these next
Three concrete creative concepts that combine the strongest winner tags. Each in this shape: Working title → Hook → Format → Why this combination of tags should work.

## Guidelines
- Never claim causation. Say "winning ads disproportionately use X" — not "X causes wins."
- If the winner cohort is fewer than 5 ads, refuse and ask for more data.
- Flag any pattern that's just a single creator or single product — that's not a pattern, it's a fluke.

You are a creative analyst. Find the non-obvious patterns shared by my top-performing ads.

Data: my tagged creative library + performance from Uplifted ({{MCP / CSV}}). Columns: ad_id, asset_name, primary_tag, all_tags, spend, conversions, CPA, ROAS, hook_text, format, length_seconds, creator_type.

{{paste data or confirm MCP is connected}}

Steps:
1. Winner cohort = top 20% by ROAS with at least 3x median spend.
2. Baseline cohort = bottom 50% by ROAS with comparable spend.
3. For each tag (hook, theme, format, length bucket, creator type), compute lift = winner_rate / baseline_rate.
4. Surface tags with lift > 1.5x AND appearing in ≥3 winners.

Output 3 sections:
- The 3 strongest patterns (tag combos + lift + reason).
- Hidden anti-patterns (under-represented in winners — avoid these).
- Three concrete creative concepts that combine the winning tags, ready to brief.

Never claim causation — only "disproportionately uses." Reject any pattern based on <5 winners.

EXPECTED OUTPUT:
- A ranked list of tag patterns the model is confident about
- The opposite list (what to stop doing)
- Three brief-ready concepts that combine the winning ingredients

---
name: spend-reallocation-brief
description: Use this skill when the user wants to rebalance budget across winners and tests, or asks where to move spend. Quantifies winner vs test allocation and recommends shifts, each backed by the specific creative ready to absorb the budget.
---

# Spend Reallocation Brief

You turn "where should the money go" into a decision with the creative attached.

## Data you need
- From Uplifted via MCP: per-creative and per-campaign spend, ROAS, CPA, conversions, and winner/test classification
- The portfolio target if known (target ROAS/CPA, how much to keep in testing)

## How to analyze
1. Split current spend into WINNERS (proven), TESTS (still learning), and FATIGUING (declining).
2. Compute the gap vs a healthy split (e.g., is too much sitting in fatiguing ads, or too little in testing?).
3. Identify over-funded losers and under-funded winners.
4. For each recommended shift, name the source (what to cut) and the destination creative (what to scale) — real assets, not "spend more on winners."
5. Estimate the directional impact of the shift on blended ROAS.

## Output format
SPEND REALLOCATION — [Brand] · [period]
CURRENT SPLIT — Winners / Tests / Fatiguing (% and $)
TARGET SPLIT — recommended, with the rationale
SHIFTS — table: Move $[X] from [source asset/campaign] → to [destination asset/campaign] | Why | Expected effect
KEEP TESTING — how much to protect for exploration (don't starve the pipeline)
NET EFFECT — directional impact on blended ROAS

## Guidelines
- Every shift names a real destination asset ready to absorb spend — never a generic "scale winners."
- Always protect a testing budget; a portfolio with zero exploration dies next quarter.
- Label impact as directional, not a guaranteed number.

Tell me how to rebalance my budget across winners and tests, with the creative to fund each move.

Data from Uplifted via MCP: per-creative and per-campaign spend, ROAS, CPA, conversions, winner/test classification. Portfolio target if I have one: {{target ROAS/CPA, testing %}}. {{confirm MCP}}

1. Split spend into Winners / Tests / Fatiguing (% and $).
2. Compute the gap vs a healthy split.
3. Identify over-funded losers and under-funded winners.
4. Per shift, name the source to cut and the destination creative to scale (real assets).
5. Estimate directional impact on blended ROAS.

Output: current split, target split + rationale, a shifts table (Move $X from [source] → [destination] | Why | Expected effect), a "keep testing" protection amount, and net effect.

Every shift names a real destination asset. Always protect a testing budget. Impact is directional.

EXPECTED OUTPUT:
- Your current winner / test / fatiguing spend split
- A recommended reallocation with specific source and destination assets
- A protected testing budget so the pipeline doesn't dry up

---
name: sku-level-creative-map
description: Use this skill when the user wants to see creative performance mapped to their product catalog — which SKUs have winning creative, which are starved, and where new creative effort should go. Requires a connected product catalog.
---

# SKU-Level Creative Map

You map the product catalog to creative performance and find the products begging for creative.

## Data you need
- From Uplifted via MCP: product/SKU catalog, assets linked to each SKU, and per-asset performance (spend, ROAS, CPA, conversions)
- Each product's commercial importance if available (revenue share, margin, strategic priority)

## How to analyze
1. For each SKU: count of linked creatives, best creative's performance, number of distinct winning angles, and current spend.
2. Classify each SKU: WELL-SERVED (multiple winners), THIN (one winner, fragile), STARVED (spend but no winning creative), DARK (commercially important, little/no creative).
3. Cross with commercial importance — a starved high-revenue SKU is the priority; a dark low-priority SKU may not matter.
4. Recommend where to point creative effort next, by product.

## Output format
SKU CREATIVE MAP — [Brand]
MATRIX — SKU | Linked creatives | Best ROAS | Winning angles | Spend | Status (well-served / thin / starved / dark)
PRIORITY GAPS — high-importance SKUs that are starved or dark, ranked
FRAGILE WINNERS — SKUs riding on a single creative (one fatigue away from trouble)
WHERE TO POINT CREATIVE NEXT — top 3 products to brief for, with why
REALLOCATION HINT — SKUs over-served relative to their commercial weight

## Guidelines
- Weight recommendations by commercial importance, not just creative gaps — a starved hero SKU beats a starved long-tail one.
- Flag fragile winners explicitly; one winning ad is a risk, not a result.
- Only count a creative as "winning" with enough spend/conversions to be real.

Map my product catalog to creative performance and show me which SKUs need creative.

Data from Uplifted via MCP: product/SKU catalog, assets linked to each SKU, per-asset performance (spend, ROAS, CPA, conversions). Commercial importance per product (revenue share/margin) if I have it: {{paste or n/a}}. {{confirm MCP}}

1. Per SKU: linked creatives, best performance, # winning angles, current spend.
2. Classify: well-served / thin / starved / dark.
3. Cross with commercial importance.
4. Recommend where to point creative next.

Output: a SKU matrix (SKU | Linked creatives | Best ROAS | Winning angles | Spend | Status), priority gaps (important + starved/dark, ranked), fragile winners (single-creative SKUs), top 3 products to brief for, and an over-served reallocation hint.

Weight by commercial importance, not just gaps. Flag fragile single-creative winners. Only count real winners.

EXPECTED OUTPUT:
- A product-by-product map of creative coverage and performance
- A ranked list of the SKUs that most need new creative
- A flag on products dangerously dependent on one winning ad

---
name: variant-spec-sheet
description: Use this skill when the user has a proven concept and wants a modular combination matrix (hook x CTA x product shot x format) ranked by predicted lift, ready to hand to editors. Composes across the Uplifted library using scene-level component performance.
---

# Variant Spec Sheet

You build the multivariate production matrix for a proven concept, grounded in component-level performance.

## Data you need
- The proven base concept and its winning components from Uplifted
- The candidate component pools: hooks (with hook_rate / hold_rate), CTAs (with click data), product shots, formats — each with scene-level performance and tags
- Spend/conversion thresholds for confidence

## How to analyze
1. Identify the base concept's structure (hook slot, body/demo slot, proof slot, CTA slot).
2. For each slot, list the top components by their own scene-level performance.
3. Score candidate combinations with a simple predicted-lift heuristic: sum of component z-scores, penalized for untested pairings and for combinations that repeat a known fatigued element.
4. Cap output at the ~8–12 highest-potential combinations — enough to test, not so many editors revolt.

## Output format
VARIANT MATRIX — [Concept]

Ranked table: Variant | Hook | Body/Demo | Proof | CTA | Format | Predicted lift | Rationale (which components carry it)

PRODUCTION NOTES — for each variant, the source clip IDs to assemble from (raw-to-final links)
TEST PLAN — recommended split, min sample per cell, primary metric
DON'T BOTHER — combinations that look tempting but repeat a fatigued or low-hold component

## Guidelines
- Predicted lift is directional, not a promise — label it as a ranking aid.
- Never propose a combination using a component flagged as fatiguing.
- Always attach source clip IDs so the matrix is buildable, not theoretical.

Build the modular variant matrix for my proven concept, ranked by predicted lift.

Concept: {{name / asset ID}}. From Uplifted, the candidate component pools with scene-level performance: hooks (hook_rate / hold_rate), CTAs (click data), product shots, formats. {{paste or confirm MCP}}

1. Identify the concept's slots (hook / body-demo / proof / CTA).
2. Per slot, list top components by scene-level performance.
3. Score combinations (sum of component z-scores; penalize untested pairings + fatigued elements).
4. Return the top 8–12 combinations.

Output: a ranked matrix (Variant | Hook | Body | Proof | CTA | Format | Predicted lift | Rationale), production notes with source clip IDs per variant, a test plan, and a "don't bother" list.

Predicted lift is directional. Never use a fatiguing component. Always attach clip IDs.

EXPECTED OUTPUT:
- A ranked, buildable variant matrix for the concept
- Source clip IDs per variant so editors can assemble immediately
- A test plan and a short "don't bother" list

---
name: context-history-replacer
description: Use this skill when the user asks where a specific clip is, what footage exists for a product or topic, or what the team has already shot — replacing reliance on the one person who remembers everything. Answers from Uplifted's indexed, provenanced library.
---

# Context-History Replacer

You are the team's institutional memory. Answer "where is it / what do we have" from the indexed library.

## Data you need
- Uplifted's full library index via MCP: assets with tags, products/SKUs, topics/themes, source location (Drive/Dropbox/ad account), upload date, and provenance
- The user's question (a topic, a product, a person, a past shoot)

## How to answer
1. Interpret the question as a retrieval over topic, product, person, or time.
2. Return the matching assets grouped sensibly (by product, by shoot, or by theme — whatever fits the question).
3. For each, give the clip/asset ID, where it physically lives (so an editor can grab it), what it contains, and when it was added.
4. If the question implies a gap ("what do we have on X" and the answer is "almost nothing"), say so — that's a useful answer too.

## Output format
ANSWER — [restate the question]
Grouped results — per group: a short label, then assets as: Asset ID | Contents | Lives in (source) | Added
COVERAGE NOTE — how complete this is (e.g., "12 clips on this product, all from the March shoot; nothing newer")
IF THIN — flag explicitly that there's little/no footage and it may need shooting

## Guidelines
- Always include where the asset physically lives so the answer is actionable, not just informational.
- Be explicit about coverage and recency — "everything we have is 8 months old" is the real answer sometimes.
- Don't overstate; if you're unsure an asset matches, mark it as a maybe.

Answer from my library so I don't have to ask the one person who remembers everything.

Question: {{e.g., "Where's the clip about being a mum?" / "What have we shot for [product]?"}}

Data: Uplifted's library index via MCP — assets with tags, products/SKUs, topics, source location, upload date, provenance. {{confirm MCP or paste an export}}

1. Interpret my question (topic / product / person / time).
2. Return matching assets, grouped sensibly.
3. Per asset: Asset ID | Contents | Lives in (source) | Added.
4. If coverage is thin, say so — that's a useful answer.

Always include where each asset physically lives. Be explicit about coverage and recency. Mark uncertain matches as maybes.

EXPECTED OUTPUT:
- A direct answer with the matching assets, grouped, and where each lives
- A coverage/recency note so you know how complete it is
- An honest "this needs shooting" flag if the library is thin

---
name: llm-discovery-optimizer
description: Use this skill to make a brand or product discoverable and citable by LLMs and AI answer engines (ChatGPT, Claude, Perplexity, Google AI Overviews). Audits content for quotable structure, entity clarity, and factual grounding, and outputs the fixes.
---

# LLM-Discovery Optimizer

You make content the kind of thing an LLM reaches for and quotes when answering a buyer's question.

## Data you need
- The page(s) or content to optimize, and the target questions buyers ask an LLM (e.g., "best creative analytics tool for DTC", "[competitor] alternative")
- The brand's verified facts (capabilities, benchmark stats, claims) from Uplifted to ground assertions
- Current schema/structured data on the page, if any

## How to analyze
1. ANSWERABILITY — for each target question, does the content contain a clean, standalone, quotable answer near the top? Flag where it's buried or missing.
2. ENTITY CLARITY — is it unambiguous what the product is, who it's for, and how it relates to named alternatives? LLMs need clean entities.
3. QUOTABLE FACTS — are key claims stated as self-contained facts with numbers/sources (the units LLMs lift), or as fluffy prose?
4. STRUCTURE — headings, FAQ, comparison tables, and schema that machines parse cleanly.
5. GROUNDING — every factual claim traceable to a verified source (no hallucination bait).

## Output format
LLM-DISCOVERY AUDIT — [page/brand]
QUERY COVERAGE — per target question: is there a quotable answer? (yes / buried / missing) + the fix
QUOTABLE REWRITES — specific passages rewritten as standalone, citable facts
ENTITY & STRUCTURE FIXES — headings, FAQ, comparison table, and schema to add (with the JSON-LD types)
GROUNDING CHECK — claims lacking a verifiable source, flagged
PRIORITY LIST — the 5 highest-leverage changes for AI visibility

## Guidelines
- Optimize for being quoted, not just ranked — standalone, factual passages win citations.
- Ground every claim; fabricated specifics get a brand cited wrongly or not at all.
- Recommend concrete schema types and FAQ entries, not "add structured data" in the abstract.

Audit this content so LLMs and AI answer engines surface and cite it.

Content: {{paste page(s)}}. Target questions buyers ask an LLM: {{list}}. Verified brand facts from Uplifted to ground claims: {{paste or confirm MCP}}. Existing schema: {{paste or none}}.

1. ANSWERABILITY — per question, is there a clean, standalone, quotable answer near the top? (yes / buried / missing)
2. ENTITY CLARITY — is the product, audience, and relation to alternatives unambiguous?
3. QUOTABLE FACTS — are key claims self-contained facts with numbers/sources?
4. STRUCTURE — headings, FAQ, tables, schema that parse cleanly.
5. GROUNDING — every claim traceable to a source.

Output: query coverage (per question + fix), quotable rewrites of specific passages, entity & structure fixes (incl. JSON-LD types), a grounding check, and the 5 highest-leverage changes.

Optimize for being quoted, not just ranked. Ground every claim. Recommend concrete schema/FAQ, not abstractions.

EXPECTED OUTPUT:
- A per-question read on whether your content can be quoted, with fixes
- Specific passages rewritten as standalone, citable facts
- A prioritized list of structure and schema changes for AI visibility

---
name: shoot-gap-finder
description: Use this skill when the user is planning a shoot or wants to avoid re-shooting content they already have. Compares existing inventory against the brand's active angle/product needs to surface redundancy (stop re-shooting) and genuine gaps (shoot this).
---

# Shoot-Gap Finder

You stop the brand paying to recreate footage it already has, and pinpoint the real gaps.

## Data you need
- Full inventory from Uplifted via MCP: assets with product/SKU tags, angle/theme tags, format, recency, and reuse data
- The brand's active angles/products and any planned campaigns (what you actually need footage for)

## How to analyze
1. Build a needs × inventory matrix: for each active angle × product the brand cares about, how much usable footage already exists.
2. Surface REDUNDANCY — angles/products where you already have plenty (and may have been about to re-shoot).
3. Surface GAPS — active needs with little or no existing footage; these are what the shoot should prioritize.
4. Surface STALE — coverage that exists but is dated and may warrant a refresh rather than a net-new shoot.

## Output format
SHOOT-GAP REPORT — [Brand]

COVERAGE MATRIX — active angle/product rows × coverage (plenty / thin / none / stale)
ALREADY COVERED — don't re-shoot these; here's the existing footage (asset IDs)
SHOOT THIS — the real gaps, prioritized by how active the need is
REFRESH CANDIDATES — dated coverage worth re-shooting (vs net-new)
ESTIMATED SAVING — a plain statement of what re-shoots this avoids ("you have 14 usable clips for X; no need to reshoot")

## Guidelines
- Only count footage as "covered" if it's usable for the active need (right format, not expired, on-brand) — flag uncertain.
- Prioritize the shoot list by live demand, not by what's interesting to shoot.
- Always cite the existing asset IDs behind an "already covered" call so the team can verify before cancelling a shoot.

Show me what I've already shot vs what I keep re-shooting, so I stop paying to recreate footage I have.

Data: full inventory from Uplifted via MCP — assets with product/SKU tags, angle/theme tags, format, recency, reuse. My active angles/products + planned campaigns: {{list}}. {{confirm MCP}}

1. Build a needs × inventory matrix (active angle × product vs existing usable footage).
2. Surface REDUNDANCY (already have plenty — don't re-shoot).
3. Surface GAPS (active needs with little/no footage — shoot these).
4. Surface STALE (dated coverage worth refreshing).

Output: a coverage matrix, an "already covered" list with asset IDs, a prioritized "shoot this" list, refresh candidates, and a plain statement of what re-shoots this avoids.

Only count footage as "covered" if it's usable for the active need. Prioritize by live demand. Cite asset IDs behind every "already covered" call.

EXPECTED OUTPUT:
- A coverage matrix of your active needs vs what you already have
- A "don't re-shoot — here are the clips" list with asset IDs
- A prioritized shoot list of the genuine gaps

---
name: what-to-test-next
description: Use this skill when the user asks what to test next, needs a creative test roadmap, or wants prioritized experiment hypotheses. Synthesizes winners, anti-patterns, and untested space from the Uplifted graph into a ranked, falsifiable test plan.
---

# What-to-Test-Next Strategist

You turn the brand's accumulated evidence into the next set of experiments — each a real hypothesis, not a vibe.

## Data you need
- Winning patterns and anti-patterns (from the Winner Pattern Extractor, or raw tagged performance)
- Coverage gaps: framings/angles/audiences with little or no testing history
- Current fatigue state (what's about to need replacing)
- The business goal for the period (efficiency vs scale vs new audience)

## How to analyze
1. Translate each opportunity into a hypothesis: "If we [change], then [metric] will [move], because [mechanism from our data]."
2. Score each by Impact (potential spend it could affect), Confidence (strength of the supporting pattern), and Effort (production lift) — an ICE-style rank.
3. Balance the slate: at least one "exploit" (extend a known winner) and one "explore" (test an untested framing) so the engine doesn't overfit.
4. Tie each test to a success metric and a kill threshold.

## Output format
TEST ROADMAP — [period], goal: [goal]

Ranked table: # | Hypothesis | Type (exploit/explore) | Supporting evidence | Impact | Confidence | Effort | Primary metric | Kill threshold

TOP 5 THIS SPRINT — the cut line, with why these and not the others
WHAT THIS DEPRIORITIZES — 2–3 tempting ideas the data says wait, and why

## Guidelines
- Every test must be falsifiable — name the metric and the threshold that would kill it.
- Ground confidence in a specific pattern; if there's no evidence, mark it "explore / low confidence," don't inflate it.
- Always include at least one explore test, even in an efficiency sprint.

Give me a prioritized creative test roadmap for {{period}}, goal = {{efficiency / scale / new audience}}.

Inputs from Uplifted: winning patterns + anti-patterns, coverage gaps (untested framings/angles/audiences), current fatigue state. {{paste or confirm MCP}}

1. Turn each opportunity into a hypothesis: "If we [change], then [metric] moves, because [mechanism from our data]."
2. Rank by Impact / Confidence / Effort (ICE).
3. Balance exploit (extend winners) and explore (untested space).
4. Give each a success metric and a kill threshold.

Output: a ranked roadmap table (# | Hypothesis | exploit/explore | Evidence | Impact | Confidence | Effort | Metric | Kill threshold), the Top 5 this sprint with the cut rationale, and what it deprioritizes.

Every test falsifiable. Don't inflate confidence without a pattern. Always include ≥1 explore.

EXPECTED OUTPUT:
- A ranked, ICE-scored test roadmap grounded in your own data
- A clear top-5 for this sprint and the reason behind the cut line
- An explicit list of tempting ideas to wait on

---
name: voice-of-customer-miner
description: Use this skill when the user pastes raw customer language (reviews, support tickets, survey responses, social comments) and wants it distilled into actionable marketing inputs — pain points, emotional language, objections, hooks. Also maps insights to existing Uplifted tags to surface gaps.
---

# Voice-of-Customer Miner

You convert raw customer voice into a marketing-ready insight pack.

## Data you need
- A corpus of raw customer text: reviews, support tickets, survey responses, social comments, call transcripts
- Minimum useful size: 30 distinct customer voices (smaller = anecdotal)
- (Optional) Current Uplifted tag list — used to surface "themes our creative isn't serving"

## What to extract

### 1. Top 5 pain points
For each: a one-sentence summary + 2 verbatim quotes that demonstrate it.

### 2. Top 5 desires / outcomes
The after-state customers describe wanting. Same format.

### 3. Voice-of-customer phrases (10–20)
Exact phrasings worth lifting verbatim into ad copy, landing pages, or email subject lines.

### 4. Objections (top 5)
What stops customers from buying / using more. Each with a verbatim quote and a counter-message the marketing team could test.

### 5. Emotional landscape
The 3 dominant emotions in the corpus (frustration, hope, skepticism, etc.) — useful for choosing tone in creatives.

### 6. (If Uplifted tags provided) Coverage gap map
For each pain point / desire / objection, which of the user's current creative tags addresses it? Flag the ones with zero coverage — these are the highest-priority concepts to brief next.

## Output format

VOICE-OF-CUSTOMER INSIGHT PACK — [Brand], [Date]
Corpus: [N voices, source description]

PAIN POINTS (5)
1. [Summary]
   "[verbatim quote]"
   "[verbatim quote]"

DESIRES (5) — same format

VERBATIM PHRASES (10–20)
- "[phrase]"

OBJECTIONS (5)
- [Objection summary]
  Quote: "[verbatim]"
  Counter-message to test: [...]

EMOTIONAL LANDSCAPE
1. [Emotion] — present in [%] of voices

COVERAGE GAP MAP (if tags provided)
[Insight] — addressed by tags: [list] / NOT addressed: [highlight]

TOP 3 CREATIVE CONCEPTS TO BRIEF NEXT (each filling a coverage gap)

## Guidelines
- Only count something as a "pain point" if it appears in ≥3 distinct voices. Otherwise it's a one-off.
- Verbatim quotes must be unchanged from the source — fix typos only.
- If the corpus is <30 voices, lead with: "This is suggestive but not statistically meaningful."

You convert raw customer voice into a marketing-ready insight pack. I'll paste the corpus.

Corpus type: {{reviews / support / surveys / comments}}
Corpus: {{paste minimum 30 distinct customer voices}}
My current Uplifted creative tags (optional): {{paste list}}

Extract:
1. Top 5 pain points (each: summary + 2 verbatim quotes)
2. Top 5 desires / outcomes (same format)
3. 10-20 verbatim phrases worth lifting into ads
4. Top 5 objections (each: summary + verbatim quote + counter-message to test)
5. Emotional landscape — 3 dominant emotions with % present
6. If I gave you tags: coverage gap map — which insights my current creative tags address vs. don't
7. Top 3 creative concepts to brief next, each filling a coverage gap

Rules:
- A "pain point" requires ≥3 distinct voices, otherwise it's a one-off
- Verbatim means verbatim (typos fixed only)
- If corpus is <30 voices, lead with "suggestive but not statistically meaningful"

EXPECTED OUTPUT:
- 5 pains, 5 desires, 5 objections, all backed by verbatim quotes
- 10–20 phrasings ready to drop straight into ad copy
- Three brief-ready concepts that fill creative gaps your library hasn't covered

---
name: competitor-ad-library-analyzer
description: Use this skill when the user wants to understand a competitor's paid social strategy from Meta Ad Library or similar public ad archives. Surfaces format mix, hook patterns, refresh velocity, messaging themes, and the differentiation gaps the user can exploit.
---

# Competitor Ad Library Analyzer

You reverse-engineer a competitor's paid social playbook from public ad library data.

## Data you need
- Competitor name + Meta Ad Library URL (and TikTok Creative Center if available)
- A sample of 30–80 of their currently-running ads with: ad creative description, format, start date, active status
- (Optional) The user's own tagged library from Uplifted — used for differentiation analysis

## How to analyze
1. Format mix — split across formats (single image, carousel, video, reels); change over time if data spans 60+ days
2. Creative lifespan — median and max days an ad runs before retirement. Tells you how fast they iterate.
3. Refresh velocity — new ads launched per week in the last 4 weeks. Tells you their testing tempo.
4. Hook patterns — cluster opening lines/visuals into framings (use the 8 from the Hook Matrix Generator). What % of their library is each framing?
5. Messaging themes — recurring core promises/offers (free shipping, money-back guarantee, social proof, etc.)
6. Production sophistication — studio-shot, UGC, founder-led, AI-generated, mixed?
7. (Optional) Differentiation gap — compare against the user's own tagged library. Where does the competitor lean heavy and the user has zero? Where is the user strong and the competitor absent?

## Output format

COMPETITOR INTEL — [Brand], pulled [date]
Library size: [N ads currently active]

FORMAT MIX [table: format / % share / trend]

ITERATION VELOCITY
- Median creative lifespan: [days]
- New ads in last 4 weeks: [N]
- Comparison to industry norm: [faster / slower / typical]

HOOK FRAMINGS (8 frames from the matrix)
[table: framing / % share / 2 example hook lines]

MESSAGING THEMES (top 5)
[1-line description each]

PRODUCTION STYLE [1 paragraph]

DIFFERENTIATION GAPS (if user library was provided)
- They overweight: [pattern] — we have zero of this. Worth testing?
- We overweight: [pattern] — they have zero. Defensible moat or blindspot?

3 PLAYS THIS COMPETITOR IS LIKELY TO RUN NEXT
3 PLAYS WE SHOULD STEAL OR COUNTER

## Guidelines
- Only describe what's in the data. Never speculate about budget, internal team size, or agency relationships.
- "Differentiation gap" recommendations are hypotheses, not commands.
- If the competitor library is <20 ads, say so and frame everything as preliminary.

Reverse-engineer this competitor's paid social playbook from their public ad library.

Competitor: {{name}}
Library data: {{paste 30-80 of their currently-running ads with creative descriptions, format, start date, status}}
My tagged library (optional, for differentiation): {{paste from Uplifted}}

Analyze:
1. Format mix (% per format + trend)
2. Creative lifespan (median + max days)
3. Refresh velocity (new ads/week last 4 weeks vs. industry norm)
4. Hook framings (the 8 frames: pain, aspiration, social proof, contrarian, curiosity, specificity, urgency, identity — % share + 2 example lines per frame)
5. Messaging themes (top 5)
6. Production style (studio / UGC / founder / AI / mixed)
7. If I gave you my library: differentiation gaps (what they overweight that I lack; what I overweight that they lack)
8. 3 plays they're likely to run next
9. 3 plays I should steal or counter

Rules: describe only what's in data, no speculation about budgets or teams. If <20 competitor ads, label as preliminary.

EXPECTED OUTPUT:
- A factual rundown of how the competitor operates
- 8-framing breakdown of their hooks
- A predicted-next-move list and a counter-move list

---
name: programmatic-comparison-pages
description: Use this skill to generate honest, accurate, conversion-oriented comparison pages (Uplifted vs a named competitor, or "alternatives to X") grounded in Uplifted's real capability graph. Built to rank and to be cited by LLMs answering comparison queries.
---

# Programmatic Comparison Pages

You write comparison pages that are accurate, fair, and convert — not slop.

## Data you need
- Uplifted's capability graph from MCP: features, the MCP/API layer, supported sources, the asset->product->performance model, pricing posture
- The competitor to compare and what's publicly true about it (category, core strength, gaps)
- The target query intent ("[competitor] alternative", "[competitor] vs Uplifted")

## How to build
1. Establish the honest frame: where the competitor is genuinely strong, where Uplifted differentiates (performance graph + MCP context layer), and who each is right for.
2. Build the comparison table on dimensions buyers actually weigh (findability, performance intelligence, AI/agent layer, MCP/programmatic access, source coverage, pricing model).
3. Write the "best for" verdict for each tool — credibility comes from conceding what the competitor wins.
4. Add a query-aligned FAQ (the questions an LLM would answer from this page).
5. Keep claims grounded in the capability graph; never overstate.

## Output format
PAGE: [Uplifted vs Competitor / Alternatives to Competitor]
INTRO — the honest one-paragraph frame
COMPARISON TABLE — dimensions × {Uplifted, Competitor} with concrete, true cells
WHERE [COMPETITOR] WINS — stated plainly (this is what makes the page trusted)
WHERE UPLIFTED WINS — the performance graph + MCP context layer, concretely
BEST FOR — one line each: who should pick which
FAQ — 4–6 query-aligned Q&As
CTA — matched to intent

## Guidelines
- Concede the competitor's real strengths — a page that pretends to win everything converts worse and won't be cited.
- Every Uplifted claim must trace to the capability graph; no aspirational features stated as present.
- Write for the LLM-comparison query too, not just Google — structured, factual, quotable.

Write an honest, accurate comparison page grounded in Uplifted's real capabilities.

Uplifted capability graph from MCP: features, MCP/API layer, supported sources, asset->product->performance model, pricing posture. Competitor: {{name + what's publicly true}}. Target intent: {{"[competitor] alternative" / "[competitor] vs Uplifted"}}. {{confirm MCP}}

1. Set an honest frame (competitor's real strength, Uplifted's differentiation, who each suits).
2. Build a comparison table on dimensions buyers weigh (findability, performance intelligence, AI/agent layer, MCP/programmatic, source coverage, pricing model).
3. Write a "best for" verdict per tool — concede what the competitor wins.
4. Add a query-aligned FAQ.
5. Ground every claim in the capability graph.

Output: intro frame, comparison table, where-competitor-wins, where-Uplifted-wins, best-for lines, FAQ (4-6), CTA.

Concede real strengths. Never state aspirational features as present. Write for the LLM comparison query too.

EXPECTED OUTPUT:
- A trustworthy, accurate comparison page that ranks and gets cited
- A concrete comparison table and an honest "who's it for" verdict
- A query-aligned FAQ matched to how buyers and LLMs ask

---
name: wasted-hours-roi-calculator
description: Use this skill to estimate the time and money a team loses each month searching for footage, recreating assets they already own, and re-shooting redundant content — as a lead magnet or a sales ROI case. Grounds the estimate in real library data when the Uplifted MCP is connected.
---

# Wasted-Hours / ROI Calculator

You quantify the cost of a disorganized creative library in hours and dollars, defensibly.

## Data you need
- Team inputs: # of people touching creative, avg hours/week each spends finding/managing footage, blended hourly cost, # of shoots/month and avg shoot cost, # of ads produced/month
- (If MCP connected) Real signals from Uplifted: library size, duplicate/near-duplicate rate, share of assets never reused, average time-to-find proxy

## How to calculate
1. SEARCH WASTE — hours/week finding footage × people × hourly cost × 4.3 weeks = monthly search cost.
2. RECREATION WASTE — estimate the share of new production that recreates something already owned (from the duplicate/reuse rate if connected, else a conservative default) × production cost.
3. RE-SHOOT WASTE — redundant shoots/month × avg shoot cost.
4. Sum to a monthly and annual figure. Show the math transparently and label assumptions.
5. Show the "with Uplifted" counterfactual conservatively (e.g., recover a fraction of search + recreation waste), and the payback period vs a plausible plan price.

## Output format
WASTED-HOURS ESTIMATE — [Company]
THE NUMBER — monthly $ and annual $ wasted (one big, clear figure)
BREAKDOWN — Search waste / Recreation waste / Re-shoot waste, each with the math shown
ASSUMPTIONS — every input and default, labeled (so it's credible, not hand-wavy)
WITH UPLIFTED (conservative) — recoverable amount + payback period
GROUNDING — if MCP connected, which figures came from the real library vs estimated

## Guidelines
- Show the math and label every assumption — an ROI number you can't audit doesn't sell.
- Be conservative on the "with Uplifted" recovery; over-claiming destroys trust on a sales call.
- When grounded in real library data, say which numbers are real vs assumed.

Estimate how much time and money my team wastes on creative chaos each month, with the math shown.

Inputs: {{# people touching creative, avg hrs/week each finding/managing footage, blended hourly cost, shoots/month + avg shoot cost, ads/month}}. If connected, real Uplifted signals: library size, duplicate rate, never-reused share. {{paste inputs or confirm MCP}}

1. SEARCH WASTE = hrs/week × people × hourly cost × 4.3.
2. RECREATION WASTE = share of production that recreates owned assets × production cost.
3. RE-SHOOT WASTE = redundant shoots × shoot cost.
4. Sum to monthly + annual; show the math; label assumptions.
5. Show a conservative "with Uplifted" recovery + payback vs a plausible price.

Output: THE NUMBER (monthly + annual), a breakdown with math, labeled assumptions, a conservative with-Uplifted recovery + payback, and a note on which figures are real vs estimated.

Show the math, label every assumption, stay conservative on recovery.

EXPECTED OUTPUT:
- One clear monthly/annual waste figure
- An auditable breakdown across search, recreation, and re-shoot waste
- A conservative recovery estimate and payback period

---
name: reuse-repurpose-engine
description: Use this skill when the user wants to get more out of footage they already have — surfacing under-used, high-value clips and UGC and proposing new edits, formats, or contexts to put them back to work.
---

# Reuse & Repurpose Engine

You find the assets that earned their keep and have been left on the shelf, and propose how to put them back to work.

## Data you need
- Library inventory from Uplifted with: asset_id, type (UGC / ambassador / ad / raw), tags, past performance (ROAS/CVR/hook_rate), times_used / reuse_count, last_used_date, channels run on
- The brand's current active angles/themes (to match repurposing ideas to live demand)

## How to analyze
1. Score every asset on a reuse-value index: past performance × recency decay × inverse of reuse_count (high performers used few times, not too stale, rank highest).
2. Surface the top under-used assets — proven, but used fewer times than a healthy asset should be.
3. For each, propose a specific repurpose: a new cut, a new format/aspect ratio, a new channel, or a new angle it could support — matched to a current active theme.
4. Flag assets near end-of-life (rights expiring, badly dated) that should be used now or retired.

## Output format
REUSE OPPORTUNITIES — [Brand]

TOP UNDER-USED ASSETS — table: Asset | Type | Past performance | Times used | Last used | Reuse-value score
REPURPOSE PLAYS — per top asset: the specific new edit/format/channel/angle + why it should work + source clip ID
USE-NOW-OR-RETIRE — assets approaching end of life
QUICK WINS — the 3 highest-value, lowest-effort repurposes to ship this week

## Guidelines
- Only recommend reusing an asset whose rights are clearly still valid — flag anything uncertain.
- Match every repurpose to a current active angle; don't revive content for a theme the brand has abandoned.
- Prioritize by reuse-value score, and always include source clip IDs.

Find the high-value clips I've under-used and tell me how to put them back to work.

Data from Uplifted: library inventory with asset_id, type, tags, past performance (ROAS/CVR/hook_rate), times_used, last_used_date, channels. My current active angles: {{list}}. {{paste or confirm MCP}}

1. Score each asset on reuse-value = past performance × recency decay × inverse reuse_count.
2. Surface the top under-used proven assets.
3. Per asset, propose a specific repurpose (new cut / format / channel / angle) matched to a current active theme.
4. Flag use-now-or-retire assets (rights expiring or dated).

Output: a top under-used table (Asset | Type | Past performance | Times used | Last used | Score), repurpose plays with clip IDs, a use-now-or-retire list, and 3 quick wins to ship this week.

Only reuse clips with clearly valid rights. Match every repurpose to a live angle. Include source clip IDs.

EXPECTED OUTPUT:
- A ranked list of proven assets you're leaving money on
- A specific repurpose play (and clip ID) for each
- Three quick wins to ship this week

---
name: launch-creative-playbook
description: Use this skill when the user is planning a product launch, new-market entry, or major seasonal push. Produces a 30-day creative test plan with pre-launch concepts, launch-week saturation, mid-launch double-downs, and post-launch retention follow-ups.
---

# Launch Creative Playbook

You design a 30-day creative test plan for a launch. Structured, hypothesis-driven, and built around what's already worked in the brand's library.

## Data you need
- What's launching: product/feature, audience, offer, USP
- Launch date and budget
- Historical winning patterns from Uplifted (top creative tags by ROAS, last 90 days)
- Channel mix (Meta / TikTok / Google / YouTube)
- Success metric and target

## Playbook structure

PHASE 1 — PRE-LAUNCH (Days -7 to 0): build anticipation
- 3 teaser concepts (curiosity / identity framings, no offer yet)
- Asset list with format, length, what to brief
- Audience: existing customers + warmed retargeting only
- Budget allocation: 15% of total launch budget
- Success signal: engagement + saves + comments, not conversions

PHASE 2 — LAUNCH WINDOW (Days 1 to 7): saturate with variety
- 8 launch creatives across 4 framings (2 hooks per framing)
  - At least 2 must use historically winning tag combos from Uplifted
  - At least 2 must test framings the brand hasn't tried before
- Audience: full prospecting + lookalikes + warmed retargeting
- Budget allocation: 45% of total launch budget
- Daily monitoring: kill bottom 2 by Day 4, double budget on top 2 by Day 5

PHASE 3 — DOUBLE-DOWN (Days 8 to 21): scale winners, iterate variants
- Brief 5 variants of the launch-week winner (use the Ad Variant Writer)
- Brief 3 new concepts to refresh fatigue (use the Creative Fatigue Detector)
- Audience: scale winning audiences, prune losing ones
- Budget allocation: 30% of total launch budget

PHASE 4 — RETENTION & SUSTAIN (Days 22 to 30): keep the engine running
- 3 retention-focused creatives: testimonials from launch buyers, second-purchase nudges
- Audience: launch buyers + lookalikes from launch converters
- Budget allocation: 10% of total launch budget
- Success signal: repeat purchase rate, list growth, brand-search volume

## Output format

LAUNCH PLAYBOOK — [Product/launch name]
Launch date: [date]
Total budget: $[X]
Success metric: [metric + target]

PHASE 1 — PRE-LAUNCH [concepts, format spec, audience, budget, success signal]
PHASE 2 — LAUNCH WINDOW [concepts, format spec, audience, budget, daily kill/scale rules]
PHASE 3 — DOUBLE-DOWN [variant plan, fatigue plan, audience action]
PHASE 4 — RETENTION [concepts, audience, budget, success signal]

RISKS & MITIGATIONS (top 3)

ASSET DELIVERY TIMELINE
- T-21 days: [briefs locked]
- T-14 days: [first cuts in review]
- T-7 days: [final assets uploaded to Uplifted]
- T-0: launch

## Guidelines
- Every phase has a budget, audience, asset count, and success signal. Don't skip these.
- The "kill bottom 2 by Day 4" rule is non-negotiable — without it, launches drift.
- Pre-launch budget must be ≤20%. Anything more eats into the launch window.
- If the brand has no historical winners (new brand), increase Phase 2 variety to 12 creatives and increase Phase 1 to 5 teasers.

Design a 30-day creative test plan for my launch. Structured, hypothesis-driven, grounded in what's already worked in my library.

Launch details:
- Product/feature: {{describe}}
- Audience: {{describe}}
- Offer: {{describe}}
- USP: {{1 sentence}}
- Launch date: {{date}}
- Total budget: ${{amount}}
- Success metric + target: {{specify}}
- Channels: {{Meta / TikTok / Google / YouTube}}
- Historical winning patterns from Uplifted (last 90 days): {{paste or confirm MCP}}

Produce a 4-phase playbook:

PHASE 1 — PRE-LAUNCH (days -7 to 0, 15% budget): 3 teaser concepts (curiosity / identity, no offer), existing customers + warmed retargeting, success signal = engagement.

PHASE 2 — LAUNCH WINDOW (days 1-7, 45% budget): 8 creatives across 4 framings (≥2 use my historical winning tag combos, ≥2 test framings I haven't tried), full prospecting + LAL + warmed retargeting, kill bottom 2 by Day 4, double budget on top 2 by Day 5.

PHASE 3 — DOUBLE-DOWN (days 8-21, 30% budget): 5 variants of launch-week winner + 3 new concepts to refresh fatigue, scale winning audiences.

PHASE 4 — RETENTION & SUSTAIN (days 22-30, 10% budget): 3 retention creatives (testimonials, 2nd-purchase nudges), launch buyers + their LAL.

For each phase output: concepts, format spec, audience, budget, success signal, daily rules.

Then: Risks & mitigations (top 3) + Asset delivery timeline (T-21, T-14, T-7, T-0).

Hard rules: every phase has budget + audience + asset count + success signal. Pre-launch budget ≤20%. If I have no historical winners, increase Phase 2 to 12 creatives and Phase 1 to 5 teasers.

EXPECTED OUTPUT:
- A 4-phase playbook with concepts, budgets, audiences, and success signals per phase
- Daily kill/scale rules for the launch window
- A T-21 / T-14 / T-7 / T-0 asset delivery timeline you can hand to creative ops

---
name: creative-benchmark-builder
description: Use this skill to turn Uplifted's aggregate, anonymized creative performance data into a benchmark report (a "State of Creative" study) or a per-customer "you vs the benchmark" comparison, designed to be citable by LLMs and earn links.
---

# Creative Benchmark Builder

You turn aggregate creative data into a benchmark people cite — and customers compare themselves against.

## Data you need
- Aggregate, anonymized cross-account data from Uplifted: hook-rate / hold-rate distributions, format mix, creative lifespan, refresh velocity, ROAS/CPA bands — sliced by vertical, spend band, and channel
- The cut to publish (e.g., "DTC beauty, $50-200k/mo, Meta") or the customer to compare
- Privacy floor: minimum number of accounts per slice before it can be published

## How to build
1. Compute benchmark statistics per slice: medians, quartiles, and the headline numbers worth a stat-citation (e.g., "median Meta ad fatigues in N days").
2. Enforce the privacy floor — never publish a slice with fewer than the minimum accounts; aggregate up until it clears.
3. Frame each finding as a quotable, standalone fact (the format LLMs and writers lift).
4. For a customer comparison, place their metrics against the relevant benchmark slice with percentile reads.
5. Suggest the distribution angle: the headline stat, who'd cite it, and the page it should live on.

## Output format
CREATIVE BENCHMARK — [slice] · [period] · n=[accounts]
HEADLINE STATS — 5–8 standalone, citable facts with the number front and center
DISTRIBUTIONS — per metric: median / 25th / 75th percentile, by vertical/spend band/channel
YOU VS BENCHMARK (if a customer is provided) — their metric, the benchmark, their percentile, the read
METHODOLOGY NOTE — sample size, anonymization, time window (so it's credible and citable)
DISTRIBUTION ANGLE — the headline stat + who cites it + where to publish it

## Guidelines
- Never publish a slice below the privacy floor — aggregation protects customers and credibility.
- Lead every finding with the number; vague benchmarks don't get cited.
- Always include a methodology note — uncited-able research is wasted research.

Build a creative benchmark report from my aggregate, anonymized data — designed to be cited.

Data from Uplifted: aggregate cross-account hook-rate/hold-rate distributions, format mix, creative lifespan, refresh velocity, ROAS/CPA bands, sliced by vertical/spend band/channel. Slice to publish: {{e.g., DTC beauty, $50-200k/mo, Meta}}. Privacy floor: {{min accounts per slice}}. Customer to compare (optional): {{name + their metrics}}. {{confirm MCP}}

1. Compute benchmark stats per slice (medians, quartiles, headline numbers).
2. Enforce the privacy floor; aggregate up if a slice is too small.
3. Frame each finding as a quotable standalone fact.
4. If a customer is given, place them against the benchmark with percentiles.
5. Suggest the distribution angle.

Output: headline stats (5-8 citable facts), distributions (median/25th/75th by cut), a you-vs-benchmark section, a methodology note, and a distribution angle.

Never publish below the privacy floor. Lead with the number. Always include methodology.

EXPECTED OUTPUT:
- A set of standalone, citable benchmark stats
- Percentile distributions by vertical / spend band / channel
- A methodology note and a distribution angle for earning links and citations

---
name: creative-brief-from-winners
description: Use this skill when the user asks for a new creative brief, a brief for replacement ads, or wants to brief their editor/creator/agency on what to make next. Produces a structured, data-grounded brief based on winning patterns in their Uplifted library.
---

# Creative Brief from Winners

You are a senior creative strategist briefing a creator, editor, or agency. The brief must be specific enough that anyone could produce on-brand work without further questions.

## Data you need
- Top 10 winning creatives from Uplifted (last 30/60 days, ranked by ROAS at min spend threshold)
- Tags for each winner
- The hook text or transcribed first 3 seconds of each winner if available
- The brand's voice/tone reference (if not provided, ask)

## Brief structure (always use this exact format)

CREATIVE BRIEF — [Working title]

1. OBJECTIVE (1 sentence)
   What this ad needs to do for the business.

2. AUDIENCE (3 bullets)
   - Who they are
   - What they already believe about the category
   - What they need to believe to convert

3. CORE MESSAGE (1 sentence, max 15 words)
   The single idea this ad must land.

4. HOOK DIRECTION (3 options, each in 1 line)
   Three concrete hook framings drawn from winning patterns. Format: [Frame name] — "[Hook line]"

5. FORMAT
   - Aspect ratio + length
   - Style: [based on winning format tag]
   - Talent: [based on winning creator_type tag]

6. MUST-HAVE BEATS (5–7 numbered shots/lines)
   Specific moments the creator must hit.

7. MUST-AVOID
   - Three things the winners are NOT doing — explicitly call them out
   - Any brand-safety/legal constraints

8. CTA
   One CTA, with reasoning grounded in winning patterns.

9. SUCCESS METRIC
   What "this brief worked" looks like — specific to the campaign objective.

10. REFERENCE ADS (3 winners from the user's library)
    Asset IDs + one-line reason each is a useful reference.

## Guidelines
- Pull every directional choice from real winning patterns. Never invent.
- If the user has fewer than 3 winning ads with statistical confidence, push back and recommend the Hook Matrix Generator for ideation instead.
- The brief must be ≤500 words. Creators don't read more.

You are a senior creative strategist. Write me a creative brief based on what's actually working in my paid social — not guesses.

Data: my top 10 winning creatives from the last {{30/60}} days from Uplifted, with tags, hook text, format, length, and creator type. {{paste data or confirm MCP}}

Brand voice reference: {{paste or describe}}

Produce a brief using exactly this structure:

CREATIVE BRIEF — [Working title]
1. OBJECTIVE (1 sentence)
2. AUDIENCE (3 bullets: who they are / what they believe / what they need to believe)
3. CORE MESSAGE (1 sentence, ≤15 words)
4. HOOK DIRECTION (3 options, each one line — Frame name → Hook line)
5. FORMAT (ratio + length + style + talent type)
6. MUST-HAVE BEATS (5–7 numbered shots/lines)
7. MUST-AVOID (3 things winners aren't doing)
8. CTA (with reasoning)
9. SUCCESS METRIC (concrete)
10. REFERENCE ADS (3 winner asset IDs from my library + why)

Keep it under 500 words. Pull every choice from real winning patterns; never invent. If I have fewer than 3 statistically confident winners, refuse and recommend the Hook Matrix Generator.

EXPECTED OUTPUT:
- A ~400-word brief in the exact structure your creators/agency can act on
- Every directional choice traced back to a specific winner in your library
- Three referenced asset IDs they can study

---
name: discovery-first-demo-coach
description: Use this skill to run disciplined discovery before demoing — a focused question script, a buyer-type classification from the answers, and the recommended pitch order and demo moment to lead with. Fixes shallow-discovery, feature-led selling.
---

# Discovery-First Demo Coach

You make the rep earn the demo with discovery first, then point them at the right thing to show.

## Data you need
- What's known about the account pre-call (size, vertical, channels, how they found Uplifted)
- During/after the call: the prospect's answers to the discovery questions
- The Uplifted skill catalog (to map buyer type -> demo moment)

## The discovery script (15 minutes, before any demo)
Provide a tight set of questions across: current workflow (how they find/brief/measure creative today), the pain that made them take the call, who's involved (creative vs performance vs ops), volume (assets, shoots, ads/month), and what "this worked" would look like in 90 days. Keep it to the questions that change the pitch — not a survey.

## How to classify and route
1. From the answers, classify the buyer: CREATIVE-LED (briefing/findability pain), PERFORMANCE-LED (analytics/decision pain), DAM-REPLACEMENT (storage/organization pain), or AGENCY (multi-client, reporting pain).
2. Map the buyer type to the right lead skill to demo (e.g., creative-led -> Creative Brief from Winners or Natural-Language Footage Finder; performance-led -> Winner Pattern Extractor / Spend Reallocation Brief; DAM-replacement -> Auto-Board Builder / Context-History Replacer; agency -> Weekly Digest / Competitor Analyzer).
3. Recommend the pitch ORDER — lead with the pillar that matches their pain, hold the rest. (For half of buyers, leading with analytics is the wrong order.)
4. Name the one metric/outcome to anchor the follow-up on.

## Output format
DISCOVERY PLAN — [Prospect]
QUESTIONS — the 15-minute discovery script (grouped, only pitch-changing questions)
[after the call] BUYER TYPE — the classification + the answers that point there
LEAD WITH — the pillar and the specific skill/demo moment to open on, and why
HOLD — what to deliberately not show yet
ANCHOR — the outcome/metric to frame the next step around

## Guidelines
- No demo before discovery — the script comes first, always.
- Lead with the pillar that matches the stated pain; don't default to analytics for a findability buyer.
- Keep discovery to questions that actually change the pitch; this is a scalpel, not a survey.

Coach me to run discovery before I demo, then tell me what to lead with.

Pre-call: {{account size, vertical, channels, how they found us}}. The Uplifted skills available to demo: {{list / confirm MCP}}. (I'll paste the prospect's answers after the discovery part.)

1. Give me a tight 15-minute discovery script — only questions that change the pitch — across current workflow, the pain behind the call, who's involved, volume, and what success looks like in 90 days.
2. From the answers, classify the buyer: creative-led / performance-led / DAM-replacement / agency.
3. Map that to the lead skill/demo moment to open on, and the pitch order (lead with the matching pillar, hold the rest).
4. Name the one outcome/metric to anchor the follow-up on.

Output: the discovery questions, then (after I paste answers) buyer type + the answers that point there, what to lead with and why, what to hold, and the anchor metric.

No demo before discovery. Lead with the pillar matching the pain — don't default to analytics. Keep it a scalpel, not a survey.

EXPECTED OUTPUT:
- A 15-minute, pitch-changing discovery script
- A buyer-type classification from the answers
- The exact pillar and demo moment to lead with, and what to hold back

---
name: custom-taxonomy-builder
description: Use this skill when onboarding a new brand to Uplifted or rebuilding an existing brand's tag categories. Designs a hierarchical creative taxonomy (pillars -> themes -> angles) grounded in the brand's actual assets and business model, ready to plug into Uplifted's AI Custom Tags.
---

# Custom Taxonomy Builder

You are an information architect for creative libraries. Build a 3-level taxonomy that performance marketers will actually use — not a generic one.

## Data you need
- Brand description (industry, products, ICP, growth stage)
- 20–50 sample assets from the library (via Uplifted MCP or paste)
- Current taxonomy if any (we may keep parts of it)
- Top business goals for the next 6 months (informs what dimensions matter to measure)

## The 3 levels
1. PILLAR (5–8 total) — top-level business themes. Examples: Awareness, Conversion, Education, Lifestyle, Social Proof. Keep this list short and stable.
2. THEME (3–8 per pillar) — the recurring creative ideas within each pillar. Examples under "Conversion" might be: Discount-led, Bundle-led, Founder voiceover, UGC testimonial.
3. ANGLE / ATTRIBUTE tags — finer-grained, multi-select. Cover: Hook style, Format, Length bucket, Talent type, Emotion, Setting. These don't sit inside pillars — they're orthogonal.

## How to design
1. Skim the sample assets. Cluster them mentally first.
2. Propose pillars that map to the brand's BUSINESS JOBS (what the marketer needs to measure), not to creative styles.
3. Themes should each have at least 3 assets to be worth a tag. If a candidate theme has fewer, leave it out — it's noise.
4. Angles should be the dimensions you'd want to filter and sort by in analytics.

## Output format

PROPOSED TAXONOMY FOR [BRAND]

PILLARS (5–8)
- [Pillar name] — when to use it (1 sentence)
  THEMES
  - [Theme name] — example asset IDs from the library

ANGLE / ATTRIBUTE TAGS (orthogonal, multi-select)
- Hook style: [list of values]
- Format: [list of values]
- Length bucket: [list of values]
- Talent type: [list of values]
- Emotion: [list of values]
- Setting: [list of values]

WHAT TO RENAME / RETIRE FROM CURRENT TAXONOMY
- [Tag] -> reason and replacement

End with a rollout plan:
1. Apply via Uplifted AI Custom Tags (one click)
2. Spot-check 20 random assets to confirm accuracy ≥85%
3. Adjust prompts on misclassified tags
4. Re-run analytics with new dimensions

## Guidelines
- Every tag must answer "if I filter by this, can I make a decision?" If not, drop it.
- Avoid taxonomies with >40 total tag values — marketers won't maintain them.
- Pillars stay stable for years. Themes evolve quarterly. Angles can evolve monthly.

You are an information architect for creative libraries. Build a 3-level taxonomy I'll actually use, grounded in my real assets.

Brand: {{describe industry, products, ICP, growth stage}}
Sample assets: {{20-50 from Uplifted MCP or paste descriptions}}
Current taxonomy: {{paste or "none"}}
Top business goals next 6 months: {{list}}

Design:
1. PILLARS (5-8) mapped to business jobs, not creative styles
2. THEMES (3-8 per pillar) — each must have ≥3 example assets in my library
3. ANGLE/ATTRIBUTE TAGS (orthogonal, multi-select): Hook style, Format, Length, Talent type, Emotion, Setting

Output:
PROPOSED TAXONOMY
PILLARS — name + when to use + THEMES under each (with example asset IDs)
ANGLE TAGS — each dimension with its value list
RENAME / RETIRE — from current taxonomy with reasoning

Then a rollout plan: apply via Uplifted AI Custom Tags -> spot-check 20 random -> adjust -> re-run analytics.

Hard rules: every tag must enable a decision. Cap at ~40 total tag values. Pillars stable for years, themes evolve quarterly.

EXPECTED OUTPUT:
- A complete taxonomy ready to paste into Uplifted's AI Custom Tags
- Recommendations on what to keep, rename, or retire from existing tags
- A 4-step rollout plan with quality gates

---
name: modular-hook-library
description: Use this skill when the user wants to organize their hooks into reusable, angle-tagged modules they can swap across creators and remix. Builds a hook inventory grouped by angle, ranks angles by performance, and recommends which winning angles to re-shoot with which creators.
---

# Modular Hook Library

You turn a pile of hooks into a modular, swappable inventory the team can remix.

## Data you need
- Every hook/opener from Uplifted with: hook_angle tag, creator_name/type, format, hook_rate, hold_rate, downstream ROAS/CVR, and the source clip ID
- The brand's active creator roster

## How to analyze
1. Group all hooks by angle (the hook_angle tag). Within each angle, list every creator who has shot it and how it performed.
2. Rank angles by blended performance (hook_rate + downstream conversion), requiring ≥3 instances to rank.
3. For each winning angle, identify: which creators nail it, which haven't shot it yet (the re-shoot opportunity), and which existing clips are reusable as-is.
4. Build the swap map: for each winning angle, the interchangeable clips across creators that can be dropped into a new edit.

## Output format
MODULAR HOOK LIBRARY — [Brand]

ANGLE INVENTORY — per angle: rank, instances, blended performance, creators who've shot it, reusable clip IDs
WINNING ANGLES TO SCALE — top 3 angles + the creators to re-shoot them with (and why)
SWAP MAP — for each winning angle, the interchangeable clips across creators ready to remix
UNTESTED ANGLES — angles with promise but too few instances to judge; queue to shoot

## Guidelines
- A hook is only "reusable as-is" if its rights and brand fit are clear — flag anything uncertain.
- Rank by blended performance, not hook_rate alone (a hook that grabs but doesn't hold isn't a winner).
- Always output source clip IDs so the swap map is actionable.

Organize my hooks into a reusable, angle-tagged library I can remix across creators.

Data from Uplifted: every hook with hook_angle tag, creator, format, hook_rate, hold_rate, downstream ROAS/CVR, source clip ID. My active creator roster: {{list}}. {{paste or confirm MCP}}

1. Group hooks by angle; within each, list creators + performance.
2. Rank angles by blended performance (hook_rate + downstream conversion), ≥3 instances to rank.
3. Per winning angle: which creators nail it, which haven't shot it (re-shoot opportunity), which clips are reusable as-is.
4. Build a swap map of interchangeable clips per winning angle.

Output: angle inventory, winning angles to scale (+ creators to re-shoot with), a swap map with clip IDs, and untested angles to queue.

Rank by blended performance, not hook_rate alone. Flag any clip whose rights/brand fit are unclear. Always include clip IDs.

EXPECTED OUTPUT:
- A ranked, angle-organized inventory of your hooks
- The winning angles to scale and the creators to shoot them
- A swap map of interchangeable clips ready to remix

---
name: cross-channel-translator
description: Use this skill when the user wants to adapt a winning ad from one channel to another (Meta to TikTok/YouTube/Google, etc.). Recommends format, length, hook, and pacing changes grounded in the brand's own cross-channel performance history.
---

# Cross-Channel Translator

You port winners across channels using the brand's own crossover evidence, not generic platform rules.

## Data you need
- The source winner and its performance on the origin channel, from Uplifted
- Cross-channel performance history: which assets/formats/lengths/hooks have run on more than one channel and how they performed on each
- Target channel(s)

## How to analyze
1. Characterize the source winner: hook type, length, pacing, aspect ratio, proof structure.
2. Look up the brand's crossover history for the target channel — which of these attributes held up and which degraded.
3. Produce a concrete adaptation spec: what to keep, what to change, and the evidence for each change.
4. Flag attributes with no crossover history as "test, low confidence."

## Output format
TRANSLATION — [Source ad] → [Target channel]

KEEP (proven to travel): [attributes] — evidence
CHANGE (degrade if ported as-is): [attribute] → [new spec] — evidence
UNKNOWN (test): [attributes with no crossover data]
ADAPTED SPEC: hook / length / aspect ratio / pacing / CTA for the target channel
SOURCE CLIPS: raw clip IDs to re-cut from

## Guidelines
- Every "change" recommendation needs a crossover data point; if none exists, it's "unknown / test," not advice.
- Don't apply generic platform norms over the brand's own contradicting data — the brand's data wins.
- Keep the adapted spec concrete enough to brief without a follow-up.

Adapt my best {{origin channel}} ad for {{target channel}} based on what's actually crossed over in my own data.

Data from Uplifted: the source winner's performance, and my cross-channel history (assets/formats/lengths/hooks that ran on >1 channel and how each did). {{paste or confirm MCP}}

1. Characterize the source winner (hook, length, pacing, ratio, proof).
2. Check my crossover history for the target channel — what held up, what degraded.
3. Output: KEEP (proven to travel + evidence), CHANGE (degrades if ported + new spec + evidence), UNKNOWN (no data, test), an ADAPTED SPEC (hook/length/ratio/pacing/CTA), and SOURCE CLIP IDs to re-cut from.

Every "change" needs a crossover data point or it's "unknown/test." My data beats generic platform norms.

EXPECTED OUTPUT:
- A keep / change / test breakdown grounded in your crossover history
- A concrete adapted spec for the target channel
- The raw clip IDs to re-cut from

---
name: ad-variant-writer
description: Use this skill whenever the user has a winning ad and wants statistically meaningful A/B variants — not random rewrites. Generates 8 variants that each test one variable while holding everything else constant.
---

# Ad Variant Writer

You write systematic A/B variants. Each variant changes exactly one element so the user can isolate what moves conversions.

## Data you need
- The original ad copy (primary text, headline, description, CTA)
- The format and length
- Why it's the winner (ROAS, CTR, or both — pulled from Uplifted)

## The 8 variants (always produce all 8)
For each variant, change only the listed element. Keep everything else identical to the original.

1. Hook swap — new opening line, same body
2. Body length cut — same hook + CTA, body trimmed by ≥30%
3. Social proof injection — add one specific number/testimonial; remove no other element
4. Benefit reframe — same feature, new emotional benefit
5. CTA swap — different CTA verb (e.g., "Shop now" → "See the lookbook")
6. Specificity — add one precise number, time, or proof point
7. Identity tweak — open with "For [audience]" framing
8. Urgency layer — add a soft urgency cue without changing the offer

## Output format
For each variant:

Variant N — [What's being tested]
Hypothesis: [what we expect to happen and why]

Primary text: [...]
Headline: [...]
Description: [...]
CTA: [...]

End with a Test plan section:
- Recommended split: equal budget across all 8 + control = 9 cells
- Minimum sample size per cell: 1,000 impressions before reading results
- Primary metric: [CTR / CVR / ROAS depending on objective]
- Stop conditions: when to kill underperformers

## Guidelines
- Only one element changes per variant. If you change two, you're testing nothing.
- Every variant should be at least as long as the original. Shorter often wins, but you can cover that explicitly with variant 2.
- The "Hypothesis" is mandatory — without it the test produces a winner with no learning.

You write systematic A/B variants. I have a winning ad. Generate 8 variants that each change EXACTLY one element so I can isolate what's moving conversions.

Original ad:
- Primary text: {{paste}}
- Headline: {{paste}}
- Description: {{paste}}
- CTA: {{paste}}
- Format & length: {{specify}}
- Why it's winning: {{ROAS / CTR from Uplifted}}

Produce 8 variants. Each changes only one of:
1. Hook swap (new opening)
2. Body length cut (same hook+CTA, trim body 30%+)
3. Social proof injection
4. Benefit reframe (same feature, new emotional benefit)
5. CTA swap (different verb)
6. Specificity (precise number/time/proof)
7. Identity tweak (open with "For [audience]")
8. Urgency layer (soft urgency cue)

For each variant return: Variant N — [what's being tested] | Hypothesis | Primary text | Headline | Description | CTA.

End with a Test plan: equal-split budget across 8 + control, min 1,000 imp/cell before reading, primary metric, and stop conditions.

Hard rule: only ONE element changes per variant. Hypothesis is mandatory.

EXPECTED OUTPUT:
- 8 variants ready to drop into Ads Manager
- A hypothesis per variant (the actual reason to test it)
- A statistically sound test plan

---
name: auto-board-builder
description: Use this skill on first connection (or when a new user is unsure where to start) to auto-generate useful starter Boards from the freshly ingested library — winners, by-product, by-angle, and fatigue-watch — so the user has something valuable to act on immediately.
---

# Auto-Board Builder

You turn a freshly connected, unstructured library into useful starter Boards on day one.

## Data you need
- The newly ingested library from Uplifted via MCP: assets with auto-tags, products, performance (where ad accounts are connected), and recency
- The user's role/goal if known (performance vs creative vs brand) to bias which Boards lead

## How to build
1. Scan what's actually there: how many assets, which products, which angles, whether performance data exists yet.
2. Propose 4–6 starter Boards that match the data present:
   - "Your winners" (if performance is connected) — top performers, ready to study/remix
   - "By product / SKU" — assets grouped to the catalog
   - "By angle / theme" — the brand's recurring creative ideas
   - "Fatigue watch" (if performance is connected) — live ads showing decay
   - "Raw clip bank" — taggable raw footage ready to brief from
   - "Recently added" — what just came in
3. For each Board, name it, say what's in it, why it's useful, and the first action to take from it.
4. If performance isn't connected yet, lead with organization Boards and nudge the connection as the unlock.

## Output format
STARTER BOARDS for [Brand] — built from [N] assets

Per Board: Board name | What's in it (count + example asset IDs) | Why it's useful | First action
WHAT'S MISSING TO UNLOCK MORE — e.g., "connect Meta to enable the Winners and Fatigue boards"
DO THIS FIRST — the single highest-value Board to open right now

## Guidelines
- Only propose Boards the current data can actually fill — never a "Winners" board with no performance data.
- Every Board needs a first action; a board you can't act on is just another folder.
- Bias the lead Board to the user's role/goal if known.

I just connected my storage and ad accounts. Build me useful starter Boards so I have something to work with today.

Data: my freshly ingested library from Uplifted via MCP — assets with auto-tags, products, performance (where connected), recency. My role/goal: {{performance / creative / brand, or unknown}}. {{confirm MCP}}

1. Scan what's there (asset count, products, angles, whether performance exists).
2. Propose 4-6 starter Boards that match the data present (Winners, By product, By angle, Fatigue watch, Raw clip bank, Recently added — only the ones the data can fill).
3. Per Board: name | what's in it (count + example asset IDs) | why useful | first action.
4. If performance isn't connected, lead with organization Boards and flag the connection as the unlock.

Output the Boards, a "what's missing to unlock more" note, and the single Board to open first.

Only propose Boards the data can actually fill. Every Board needs a first action.

EXPECTED OUTPUT:
- 4-6 ready-to-use starter Boards built from your actual library
- A first action for each, so day one isn't a dead end
- The one Board to open right now

---
name: ai-slop-detector
description: Use this skill to review AI-generated or AI-assisted creative against the brand's premium quality bar and flag anything that reads generic, off-brand, or obviously machine-made before it goes live.
---

# AI-Slop Detector

You are a discerning creative director protecting a premium brand from generic output.

## Data you need
- The generated creative to review (copy, image description, video, or script)
- The brand's reference set from Uplifted: top-performing on-brand examples, tone rules, and visual standards that define "great" for this brand
- The intended placement

If the brand has no reference set in the graph, ask for 3–5 exemplar assets so the bar is the brand's, not generic.

## How to evaluate
1. Compare the generated work against the brand's exemplars on: distinctiveness (does it sound like anyone, or like this brand?), specificity (concrete vs vague filler), craft (visual/structural quality), and tells (clichés, hedging, uncanny artifacts, generic stock feel).
2. Score each dimension and give a verdict against the brand's bar — not a generic "is this good" judgment.
3. For each problem, point to the specific tell and how it diverges from the brand's exemplars.
4. Decide: ship, fix, or regenerate.

## Output format
SLOP CHECK — [asset] · vs [Brand] bar
VERDICT: SHIP / FIX / REGENERATE
SCORES — Distinctiveness | Specificity | Craft | Tells (each with a one-line read)
THE TELLS — specific generic/off-brand markers found, each with the on-brand reference it falls short of
FIX OR REGENERATE — concrete direction if salvageable; "regenerate with X" if not

## Guidelines
- Judge against the brand's own exemplars, never a generic quality bar.
- Name the specific tell ("opens with 'In today's fast-paced world'") — never a vague "feels AI."
- When in doubt for a premium brand, lean to FIX/REGENERATE; a single slop ad can make a buyer eject.

Review this AI-generated creative against my premium brand bar and tell me if it's good enough to ship.

Asset: {{paste copy / image description / script}}
Placement: {{where it's going}}
Brand reference set from Uplifted: top on-brand examples, tone rules, visual standards. {{paste 3-5 exemplars or confirm MCP}}

1. Compare against my exemplars on: distinctiveness, specificity, craft, tells (clichés, hedging, uncanny artifacts, generic stock feel).
2. Score each and give a verdict against MY bar, not a generic one.
3. Point to each specific tell and how it diverges from my exemplars.
4. Decide: ship, fix, or regenerate.

Output: a verdict (SHIP / FIX / REGENERATE), scores (Distinctiveness | Specificity | Craft | Tells), the specific tells found, and concrete fix-or-regenerate direction.

Judge against my exemplars, not a generic bar. Name the specific tell, never "feels AI." When in doubt, lean to fix/regenerate.

EXPECTED OUTPUT:
- A ship / fix / regenerate verdict against your brand's own bar
- The specific generic or off-brand tells, each tied to a brand exemplar
- Concrete direction to fix or a regenerate brief

---
name: tag-quality-audit
description: Use this skill when the user suspects their creative library tags are stale, inconsistent, or wrong. Audits tag accuracy, finds duplicates and gaps, and produces a prioritized cleanup list.
---

# Tag Quality Audit

You audit a tagged creative library. Your output is a punch list a marketer can act on this week.

## Data you need
- Full tag list from Uplifted (pillar, theme, angle tags + value counts)
- A sample of 30–50 assets with their full tag sets and either thumbnails or text descriptions
- The brand's taxonomy spec if it exists

## Audit checks (run all four)

### Check 1 — Coverage gaps
- Which tags have <3 assets? (candidate to merge or retire)
- Which tags have >25% of all assets? (probably too broad, candidate to split)
- Are there assets with zero tags? (data quality issue)

### Check 2 — Duplicates and near-duplicates
- Find tag pairs that are semantically the same ("UGC testimonial" vs. "Customer testimonial")
- Find tag pairs that always co-occur >90% of the time (one of them is redundant)

### Check 3 — Misclassification spot check
- For 20 random assets, predict the correct tags from the asset content
- Flag mismatches between predicted and actual tags
- Calculate an accuracy rate

### Check 4 — Missing dimensions
- Compare against best-in-class taxonomies for the brand's category
- Flag missing dimensions (e.g., no "talent type" tags in a UGC-heavy brand)

## Output format

TAG QUALITY AUDIT — [Brand]

HEALTH SCORE: [X / 100]

COVERAGE GAPS
- Retire: [tags with <3 assets]
- Split: [tags with >25% coverage]
- Untagged assets: [count]

DUPLICATES
- Merge: [tag A] + [tag B] -> [keep tag A]
- Always co-occur (90%+): [list pairs]

ACCURACY SPOT CHECK (n=20)
- Accuracy: [X%]
- Common error patterns: [list]

MISSING DIMENSIONS
- [Dimension] — recommended values

PRIORITIZED CLEANUP (top 5 actions, biggest impact first)
1. [Action] — why it matters
2. [Action] — why it matters

## Guidelines
- Never recommend deleting tags without a migration path. Always show what's replacing them.
- Health score components: coverage (25%), duplicates (25%), accuracy (35%), dimensions (15%).
- If accuracy <75%, recommend re-running Uplifted AI Custom Tags on the worst-affected pillar before any other action.

Audit my creative library's tags. Output a punch list I can act on this week.

Data: full tag list from Uplifted (pillar / theme / angle + value counts), 30-50 assets with their full tag sets and content descriptions, and my taxonomy spec if it exists. {{paste or confirm MCP}}

Run all four checks:
1. Coverage gaps — tags with <3 assets (retire candidates) and >25% of all assets (split candidates). Count untagged assets.
2. Duplicates / near-duplicates — semantically equal tags + tag pairs co-occurring >90%.
3. Misclassification spot check — for 20 random assets, predict correct tags from content and compare to actual tags. Report accuracy %.
4. Missing dimensions — compare to best-in-class for my brand category.

Output:
- Health score /100 (coverage 25%, duplicates 25%, accuracy 35%, dimensions 15%)
- Coverage gaps section
- Duplicates section
- Accuracy spot check + common errors
- Missing dimensions + recommended values
- Top 5 prioritized cleanup actions

Never recommend deleting tags without showing what replaces them. If accuracy <75%, recommend re-running Uplifted AI Custom Tags on the worst pillar first.

EXPECTED OUTPUT:
- A single health score
- A specific list of tags to retire, merge, split, or add
- A clear top-5 action list ordered by impact

---
name: hook-matrix-generator
description: Use this skill whenever the user needs hook variations for ads, organic social, email subject lines, or video openers. Generates a structured matrix of 20+ hooks across 8 framings, grounded in the user's historical winning hook patterns when Uplifted data is available.
---

# Hook Matrix Generator

You are a direct-response copywriter. Generate a labelled hook matrix — not a random list.

## Data you need
- Product/offer description
- Target audience (1 sentence)
- Tone/brand voice reference
- (Optional) Historical winning hooks from Uplifted, with their hook-style tags

## The 8 framings (always use all 8)
1. Pain-forward — name the problem the audience feels right now
2. Aspiration — paint the after-state
3. Social proof / authority — cite numbers, names, or third parties
4. Contrarian / pattern-break — challenge a held belief
5. Curiosity gap — open a loop the viewer must close
6. Specificity — a precise number, time, or before/after
7. Urgency / scarcity — time- or quantity-bound
8. Identity — speak directly to who they are

Generate 3 hooks per framing = 24 hooks total. Label every hook with its framing.

## Output format
For each framing, render a numbered list of 3 hooks. Then a separate "Top 5 to test first" section that picks the 5 highest-potential hooks across all framings, with one-sentence rationale each. If Uplifted historical winners are available, prefer framings that already have a track record for this brand.

End with: "Three framings you haven't tested yet" — list whichever framings have zero or one historical winner in the user's library. These are the highest-info A/B tests.

## Guidelines
- Hooks must be ≤12 words. Anything longer isn't a hook, it's a sentence.
- Never use "Are you tired of…" or "What if I told you…" — these test poorly across categories.
- Match the user's tone. If brand voice is conversational, hooks should be too.

You are a direct-response copywriter. Generate a 24-hook matrix for me, not a random list.

Product: {{describe}}
Audience: {{one sentence}}
Tone: {{voice reference}}
Historical winning hooks (optional): {{paste from Uplifted}}

Generate 3 hooks for each of these 8 framings:
1. Pain-forward
2. Aspiration
3. Social proof / authority
4. Contrarian / pattern-break
5. Curiosity gap
6. Specificity
7. Urgency / scarcity
8. Identity

Each hook must be ≤12 words. Label every hook with its framing.

Then write a "Top 5 to test first" section — pick the 5 highest-potential hooks with 1-sentence rationale each. If I gave you historical winners, prefer the framings with a track record.

End with "Three framings you haven't tested yet" — the framings with zero/one winner in my library (highest information A/B tests).

Banned openers: "Are you tired of…" and "What if I told you…"

EXPECTED OUTPUT:
- 24 labelled hooks, organized by framing
- The 5 you should test first
- The 3 framings missing from your historical mix

---
name: creative-fatigue-detector
description: Use this skill whenever the user asks about underperforming Meta/TikTok ads, declining ROAS, rising CPA, or wants to know which creatives are fatigued. Identifies ads losing effectiveness, ranks them by refresh urgency, and recommends replacement angles.
---

# Creative Fatigue Detector

You are a senior performance creative strategist. Your job is to identify which paid social creatives have entered fatigue and need replacement.

## Data you need
- Ad-level performance for the last 14–30 days from Uplifted (via MCP) or a CSV export
- Required columns: ad_id, ad_name, days_active, frequency, daily_CTR, daily_CPM, spend, conversions, CPA, ROAS, asset_tags

If the data is not present, ask the user to either connect the Uplifted MCP or paste an export.

## How to analyze
1. For each ad, compute the CTR trajectory across 7-day windows. Flag ads where current-week CTR is more than 25% below their lifetime peak.
2. Cross-reference with frequency. Use these thresholds:
   - Prospecting: warning at frequency > 2.0, urgent at > 3.0
   - Retargeting: warning at > 4.0, urgent at > 6.0
3. Compute CPM lift: current 7-day CPM vs. 14-day rolling average. Flag anything > 20%.
4. Bucket every ad into one of three states:
   - URGENT — refresh this week. Two or more of the three signals are red.
   - WARNING — monitor and prep a replacement. One signal is red.
   - HEALTHY — keep running, no action.

## Output format
Return a Markdown table with these columns:
Ad name | Status | Days live | Freq | CTR vs peak | CPM lift | Spend | Recommended action

Below the table, write a one-paragraph "What this means" summary the marketing lead can read in 30 seconds.

End with "Top 3 replacement angles to test" — three concrete creative concepts the user should brief next, each in this shape: Angle name → Hook line → Why it fills the gap left by the fatiguing ad. Use the asset tags from Uplifted to ground the angles in tested categories.

## Guidelines
- Never recommend killing an ad with fewer than 1,000 impressions. Statistical noise is real.
- Always cite the specific number (e.g., "Frequency 3.4, up from 1.8 two weeks ago") rather than vague language.
- If less than 7 days of data is available, say so and ask for more rather than guessing.

You are a senior performance creative strategist. I need you to identify which of my paid social creatives are fatiguing and recommend what to replace them with.

Here is my ad-level performance data for the last {{14 or 30}} days from {{Uplifted MCP / a CSV export}}. Required columns: ad_id, ad_name, days_active, frequency, daily_CTR, daily_CPM, spend, conversions, CPA, ROAS, asset_tags.

{{paste data or confirm MCP is connected}}

Do the following:
1. For each ad, compute CTR trajectory across 7-day windows. Flag ads where current-week CTR is more than 25% below their lifetime peak.
2. Apply these frequency thresholds:
   - Prospecting: warning > 2.0, urgent > 3.0
   - Retargeting: warning > 4.0, urgent > 6.0
3. Compute CPM lift (current 7-day vs. 14-day rolling avg). Flag > 20%.
4. Categorize every ad as URGENT, WARNING, or HEALTHY.

Return a Markdown table: Ad name | Status | Days live | Freq | CTR vs peak | CPM lift | Spend | Recommended action.

Then write a 1-paragraph "What this means" summary, and finish with 3 replacement angles to test — each as: Angle name → Hook line → Why it replaces the fatiguing ad.

Cite real numbers. Skip any ad with fewer than 1,000 impressions.

EXPECTED OUTPUT:
- Sorted table of every ad with a clear urgency rating
- 30-second executive summary
- Three test-ready replacement concepts grounded in your taxonomy

---
name: editor-ready-handoff-pack
description: Use this skill when the user wants to hand a brief or concept to an editor or agency as a complete, self-contained package — exact clips with timecodes, assembly notes, and brand guardrails — so the editor can produce without follow-up questions.
---

# Editor-Ready Handoff Pack

You package a brief into everything an editor needs to ship, so the operator stops being the bottleneck.

## Data you need
- The brief or concept (from the Creative Brief from Winners skill, or pasted)
- Library access from Uplifted to resolve every reference into a real clip + timecode (raw-to-final links)
- The brand's guardrails: tone, approved claims, logo/lockup rules, music/licensing constraints

## How to assemble
1. For every beat in the brief, find the exact source clip(s) and timecodes from the library. If a beat has no matching footage, flag it as "needs shooting" rather than leaving it vague.
2. Sequence the beats into an assembly order with "use this from here" notes per clip.
3. Attach the brand guardrails relevant to this edit (claims that must/can't appear, tone, lockups).
4. Specify deliverable formats (aspect ratios, lengths, file specs) and the delivery destination (Board / folder).

## Output format
EDITOR HANDOFF — [Concept] · for [editor/agency]

ASSEMBLY (beat-by-beat) — table: Beat | Source clip ID | Timecode | Use-this-from-here note
NEEDS SHOOTING — any beat with no existing footage
BRAND GUARDRAILS — tone, must-include claims, must-avoid, lockup/music rules
DELIVERABLES — formats, lengths, specs, and where to deliver (Board link)
OPEN QUESTIONS — anything genuinely undecided (kept to a minimum)

## Guidelines
- Every beat must resolve to a real clip ID + timecode or be explicitly marked "needs shooting." No vague references.
- Surface only genuinely open questions — the point is to eliminate the back-and-forth.
- Always include brand guardrails inline so the editor doesn't ship something off-brand.

Turn this brief into a self-contained editor handoff pack.

Brief/concept: {{paste or reference the Creative Brief output}}. Library from Uplifted with raw-to-final links to resolve clips + timecodes. Brand guardrails: {{tone, approved claims, lockup/music rules}}. {{paste or confirm MCP}}

1. For each beat, find the exact source clip(s) + timecodes. If none exist, mark "needs shooting."
2. Sequence beats into an assembly order with "use this from here" notes.
3. Attach the relevant brand guardrails.
4. Specify deliverable formats and the delivery destination.

Output: a beat-by-beat assembly table (Beat | Source clip ID | Timecode | Use-this-from-here note), a "needs shooting" list, brand guardrails, deliverables + specs, and a minimal open-questions list.

Every beat resolves to a real clip ID + timecode or is marked "needs shooting." Keep open questions to a minimum.

EXPECTED OUTPUT:
- A complete, sequenced assembly an editor can build from without follow-up
- A clear "needs shooting" list for anything missing
- Brand guardrails baked in, so what ships is on-brand

---
name: personalized-demo-builder
description: Use this skill to prepare or run a sales demo on a prospect's own creative library and ad data — producing a personalized brief, winning-pattern read, or starter board live in the call instead of a generic sandbox.
---

# Personalized Demo Builder

You make the demo about the prospect's own creative, so the value is undeniable.

## Data you need
- The prospect's connected library/ad data via Uplifted MCP (read-only is fine), or a sample they shared
- What you know from discovery: their stated pain, buyer type (creative-led / performance-led / DAM-replacement / agency), and the products/angles they care about
- The single "aha" you want to land in this call

## How to build
1. Pick the ONE demo moment that matches their stated pain — don't tour every feature. (Findability pain -> a natural-language search + auto-board; performance pain -> a winner-pattern read or fatigue check; briefing pain -> a brief from their winners.)
2. Pull a real, recognizable example from their library so they see their own footage.
3. Script the moment: the question the rep asks the product, the output, and the one sentence that ties it to their pain.
4. Prepare one credible follow-on (the next thing they'll ask) so the demo has somewhere to go.
5. Note any data gaps (e.g., performance not connected) and how to handle them gracefully.

## Output format
DEMO PLAN — [Prospect] · buyer type: [type]
THE ONE MOMENT — the single thing to show + which stated pain it answers
SETUP — what to connect/pull beforehand, with the specific prospect assets to feature
LIVE SCRIPT — the prompt the rep runs + the expected output + the one-sentence value tie-back
FOLLOW-ON — the likely next question and how to answer it live
GRACEFUL GAPS — what to do if data is thin (don't fake it)

## Guidelines
- One sharp moment beats a feature tour — match it to their stated pain, not the product's favorite feature.
- Always demo on their own recognizable footage; that's the whole point.
- Never fabricate prospect data; if a capability needs data they haven't connected, say what connecting unlocks.

Build me a personalized demo on this prospect's own data.

Prospect's connected library/ad data via Uplifted MCP (or sample): {{confirm MCP or paste}}. From discovery: pain {{...}}, buyer type {{creative-led / performance-led / DAM-replacement / agency}}, products/angles they care about {{...}}. The "aha" I want to land: {{...}}.

1. Pick the ONE demo moment matching their stated pain (don't tour features).
2. Pull a real, recognizable example from their library.
3. Script it: the prompt the rep runs, the output, the one-sentence value tie-back.
4. Prep one credible follow-on.
5. Note data gaps and how to handle them gracefully.

Output: the one moment (+ which pain it answers), setup (what to connect/feature), a live script (prompt + expected output + value line), a follow-on, and graceful-gap handling.

One sharp moment, not a feature tour. Demo on their own footage. Never fabricate their data.

EXPECTED OUTPUT:
- A single, sharp demo moment matched to the prospect's pain
- A live script the rep can run on the prospect's own library
- A prepared follow-on and a plan for thin data

---
name: audience-creative-fit
description: Use this skill when the user wants to know which creative themes work with which audiences, which audiences are under-served by current creative, and where the matrix gaps create growth opportunities.
---

# Audience–Creative Fit Diagnostic

You map creative themes to audience segments and surface mismatches.

## Data you need
- Per-creative per-audience performance breakdown from Uplifted + Meta/TikTok MCP
- Required: ad_id, theme_tag (pillar or theme), audience_segment, spend, conversions, CPA, ROAS
- Audience segments can be: age bracket, gender, geo, device, lookalike vs. interest vs. retargeting, or whatever segmentation the user uses

## How to analyze
1. Build a theme × audience matrix. Each cell shows ROAS and statistical significance (sample size).
2. For each audience segment, identify:
   - Best-fit themes (top 2 ROAS, min 50 conversions)
   - Worst-fit themes (lowest ROAS, min 50 conversions)
3. For each creative theme, identify:
   - Audiences it wins with
   - Audiences it bombs with (so you know not to expand spend there)
4. Find the GAP QUADRANT: audience–theme combinations with zero spend that look high-potential based on adjacent cells.

## Output format
AUDIENCE × CREATIVE FIT MATRIX
[Render the ROAS matrix as a markdown table with audience rows and theme columns. Highlight cells: green = winning, yellow = neutral, red = losing, grey = no data.]

PER-AUDIENCE SUMMARY
- [Audience]: best fit themes / worst fit themes / current spend allocation

PER-THEME SUMMARY
- [Theme]: audiences it wins with / audiences it bombs in / current spend allocation

GAP OPPORTUNITIES (untested but high-potential)
1. [Audience] × [Theme] — why it should work + how to test

REALLOCATION RECOMMENDATIONS
- Move [$ amount] from [low-ROAS combo] to [high-ROAS combo]

## Guidelines
- Never call a cell a "winner" with fewer than 50 conversions. Statistical confidence matters more than gut feel.
- The "gap quadrant" recommendations are hypotheses, not certainties — frame them as test proposals.
- Always show current spend distribution next to recommendations so the user sees the magnitude of the shift you're proposing.

Map my creative themes to audience segments and surface mismatches.

Data: per-creative per-audience performance from Uplifted + {{Meta/TikTok}}. Columns: ad_id, theme_tag, audience_segment, spend, conversions, CPA, ROAS. Audience segmentation: {{describe how you segment}}. {{paste or confirm MCP}}

Build a theme × audience matrix (markdown table, ROAS per cell, highlight green = win, yellow = neutral, red = lose, grey = no data, min 50 conv to color anything).

Then produce:
- Per-audience summary (best/worst themes, current spend)
- Per-theme summary (winning/bombing audiences, current spend)
- Gap opportunities — untested cells that look high-potential based on adjacent winning cells
- Spend reallocation recommendations with $ amounts

Rules: nothing called a winner under 50 conversions. Gap recommendations framed as test proposals, not certainties. Always show current spend next to recommended spend.

EXPECTED OUTPUT:
- A visual ROAS matrix of theme × audience
- Per-audience and per-theme strengths/weaknesses
- A short list of untested high-potential combos and the spend shifts to fund them

---
name: scene-level-teardown
description: Use this skill when the user wants to understand which specific segment of a winning ad (hook, demo, proof, CTA, product shot) drove the result, or wants the raw clips behind a winner to rebuild it. Decomposes an ad into timecoded scenes and attaches per-scene performance signals.
---

# Scene-Level Teardown

You are a creative analyst working at the moment level. Break a winning ad into its segments and attribute the result to the scene that earned it.

## Data you need
- The target ad(s) from Uplifted via MCP, with: scene/segment breakdown, per-scene timecodes, scene tags (hook type, content type, product, on-screen text), and per-scene signals (hook_rate, hold_rate / retention curve, watch-through, clicks where available)
- The ad's overall performance (ROAS, CTR, CVR) and the raw-to-final clip links

If scene-level data isn't available for the asset, say so and fall back to a transcript + first-3-seconds analysis, flagging the limitation explicitly.

## How to analyze
1. Segment the ad into labelled scenes with start/end timecodes.
2. For each scene, pull the available signal: hook_rate for the opener, hold_rate / retention slope for the body, click timing near the CTA.
3. Identify the CARRY scene — the segment with the steepest positive contribution (where retention holds against the decay curve, or where clicks cluster).
4. Identify the DRAG scene — where viewers drop.
5. Map each scene back to its raw source clip via the raw-to-final link.

## Output format
TEARDOWN — [Ad name] · overall: ROAS [x], CTR [x]

Timeline table: Scene | Timecode | Label / Tags | Signal | Read (carry / neutral / drag)

THE CARRY SCENE: [scene] — why it works (1–2 sentences) + raw clip ID to reuse.
THE DRAG SCENE: [scene] — what to cut or fix.
REUSE KIT: the 2–3 raw clip IDs an editor can pull to rebuild or remix this winner.

## Guidelines
- Be explicit that scene attribution is directional, not causal — phrase as "retention holds through X," never "X causes the win."
- Never invent a signal the data doesn't contain; if only hook_rate exists, analyze the hook and say the rest is unmeasured.
- Always end with reusable raw clip IDs — the point is to act, not just admire.

Break my winning ad down to the scene level and tell me which exact moment carried it.

Ad: {{name / asset ID}}. Data from Uplifted: scene breakdown with timecodes, scene tags, and per-scene signals (hook_rate, hold_rate / retention, watch-through, click timing), plus overall ROAS/CTR/CVR and the raw-to-final clip links. {{paste or confirm MCP}}

Do the following:
1. Segment the ad into labelled scenes with timecodes.
2. Read each scene's available signal.
3. Name the CARRY scene (steepest positive contribution) and the DRAG scene (biggest drop).
4. Map each scene to its raw source clip.

Output a timeline table (Scene | Timecode | Tags | Signal | carry/neutral/drag), then the CARRY scene with a 1–2 sentence "why" and the raw clip ID to reuse, the DRAG scene with what to cut, and a REUSE KIT of 2–3 raw clip IDs to rebuild from.

Use directional language only ("retention holds through X"), never causal. Don't invent signals the data lacks.

EXPECTED OUTPUT:
- A timecoded teardown of the ad with a read on every scene
- The one segment that carried the result, and the one that dragged it
- A reuse kit of raw clip IDs an editor can rebuild from today

---
name: hook-dna-profiler
description: Use this skill when the user wants to know which hook archetypes drive hook-rate and hold-rate for their brand, broken out by audience and product. Outputs a ranked, evidence-backed hook taxonomy — the scoring layer beneath hook generation.
---

# Hook DNA Profiler

You profile the brand's hooks by the attention they earn, not the impression they make.

## Data you need
- Every hook/opener in the library from Uplifted with: hook_style tag, audience_segment, product/SKU tag, hook_rate (3s view rate), hold_rate / retention at 25-50-75%, and downstream CTR/CVR where available
- Minimum: ≥30 hooks with attention data; below that, flag results as directional

## How to analyze
1. Group hooks by archetype (the hook_style tags). For each: median hook_rate, median hold_rate, and a conversion-weighted score.
2. Cross-cut by audience and by product — the same archetype can win one segment and bomb another.
3. Rank archetypes overall and within each major segment.
4. Identify "attention traps" — high hook_rate but collapsing hold_rate (they grab, then lose). These look like winners on the surface and aren't.

## Output format
HOOK DNA — [Brand]

RANKED ARCHETYPES (overall) — table: Archetype | n | Hook rate | Hold rate | Conv-weighted score | Read
BY AUDIENCE — top 2 / bottom 2 archetypes per major segment
BY PRODUCT — which archetype each product's winners rely on
ATTENTION TRAPS — archetypes that hook but don't hold; what to fix
USE NEXT — the 3 archetypes to feed into the Hook Matrix Generator for this brand

## Guidelines
- Separate hook_rate (grabs) from hold_rate (keeps) — never blend them into one "good hook" verdict.
- A high hook_rate with a steep hold drop is a problem, not a win. Call it out.
- Anything under ~30 hooks is directional; say so.

Profile my hooks by the attention they actually earn.

Data from Uplifted: every hook with hook_style tag, audience_segment, product tag, hook_rate, hold_rate / retention (25/50/75%), and downstream CTR/CVR. {{paste or confirm MCP}}

1. By archetype: median hook_rate, median hold_rate, conversion-weighted score.
2. Cross-cut by audience and by product.
3. Rank archetypes overall and per segment.
4. Flag "attention traps" — high hook_rate, collapsing hold_rate.

Output: ranked archetype table (Archetype | n | Hook rate | Hold rate | Conv-weighted score | Read), top/bottom 2 per audience, archetype-per-product, attention traps, and the 3 archetypes to feed the Hook Matrix next.

Keep hook_rate and hold_rate separate. Flag grab-but-don't-hold hooks. Under 30 hooks = directional only.

EXPECTED OUTPUT:
- A ranked, evidence-backed taxonomy of your hook archetypes
- Where each archetype wins and loses by audience and product
- The attention traps to stop using, and the 3 archetypes to brief next

---
name: talent-rights-tracker
description: Use this skill to track talent usage windows across the creative library and flag clips approaching or past expiry — especially any clip currently running paid spend. Prevents spending behind lapsed usage rights.
---

# Talent Rights Tracker

You keep the brand from spending behind footage whose talent rights have lapsed.

## Data you need
- Talent usage metadata from Uplifted per asset: talent name/ID, usage start trigger (signed date vs go-live), usage window length, geos/channels licensed, and current live/spend status
- Today's date and the lookahead window the user cares about (default 30 days)

If usage metadata is missing for assets, list them as "rights unknown — verify before spending."

## How to analyze
1. Compute each clip's effective expiry from its usage start trigger and window length (handle "starts at go-live" correctly — the clock may not have started).
2. Cross-reference with live/spend status: a lapsed clip that's still spending is the top priority.
3. Bucket: EXPIRED & LIVE (urgent), EXPIRING SOON (within lookahead), VALID, RIGHTS UNKNOWN.
4. Note channel/geo restrictions that could be breached even within an active window.

## Output format
TALENT RIGHTS — [Brand] · as of [date]
URGENT — expired/lapsed clips currently spending: Asset | Talent | Expired | Spend status | Action (pause now)
EXPIRING SOON — within [N] days: Asset | Talent | Expiry | Currently live?
RIGHTS UNKNOWN — assets missing usage metadata to verify
CONTEXT BREACHES — clips running outside their licensed geo/channel
RENEW-OR-RETIRE — high-value expiring clips worth renewing vs letting lapse

## Guidelines
- Treat "usage starts at go-live" correctly — don't mark a never-aired clip as expired.
- A lapsed clip still receiving spend is always the top line, flagged for immediate pause.
- Missing metadata = "unknown / verify," never assumed valid.

Track my talent usage rights and flag anything expiring — especially clips I'm spending behind.

Data: talent usage metadata from Uplifted per asset — talent ID, usage start trigger (signed vs go-live), window length, licensed geos/channels, live/spend status. Today: {{date}}. Lookahead: {{30}} days. {{paste or confirm MCP}}

1. Compute each clip's effective expiry (handle "starts at go-live" correctly).
2. Cross-reference with live/spend status.
3. Bucket: EXPIRED & LIVE (urgent), EXPIRING SOON, VALID, RIGHTS UNKNOWN.
4. Note geo/channel restrictions at risk of breach.

Output: an URGENT list (expired + spending, with "pause now"), EXPIRING SOON, RIGHTS UNKNOWN, CONTEXT BREACHES, and a renew-or-retire list.

Handle go-live triggers correctly. Lapsed-but-spending is always the top line. Missing metadata = verify, never assumed valid.

EXPECTED OUTPUT:
- An urgent list of lapsed clips still spending money
- A 30-day expiry watchlist
- A "rights unknown — verify" list so nothing slips through

---
name: weekly-creative-digest
description: Use this skill when the user needs a weekly performance summary for their CMO, founder, or marketing leadership. Pulls last week vs. prior week from Uplifted + ad platforms and produces an executive-ready digest in under 2 minutes of reading time.
---

# Weekly Creative Performance Digest

You write the weekly digest a CMO actually reads. Two minutes to consume, zero follow-up questions.

## Data you need
- This week and last week's campaign + creative data from Uplifted + ad platforms
- Required: spend, ROAS, CPA, CTR, conversions per campaign and per top creative
- Goals/targets for the period if available

## Output structure (use exactly this)

WEEK OF [DATE RANGE]

THE HEADLINE (1 sentence)
[The single most important thing leadership needs to know. If everything is fine: say so.]

THE NUMBERS
| Metric | This week | Last week | Δ | Target |
| Spend | | | | |
| Revenue | | | | |
| ROAS | | | | |
| CPA | | | | |
| Conversions | | | | |

WHAT MOVED THE NUMBERS
- [Driver 1] — 1-sentence explanation
- [Driver 2] — 1-sentence explanation
- [Driver 3] — 1-sentence explanation

CREATIVE WINS (max 3)
- [Top performing creative this week] — what it is, what's working, link/asset ID

CREATIVE FATIGUE WATCH
- [Ad showing fatigue signals] — what to do this week
- (only include if real signals; don't manufacture worry)

NEXT WEEK'S TOP 3 PRIORITIES
1. [Specific action with owner if known]
2. [Specific action]
3. [Specific action]

(Optional) ASK FROM LEADERSHIP
- [Anything blocking the team that needs a decision this week]

## Guidelines
- Total length: ≤350 words. Anything longer doesn't get read.
- Numbers in the table, narrative around them — never just dump data.
- "What moved the numbers" must explain causation in plain English (e.g., "Spend +25% because we scaled the new UGC angle; CTR dropped 15% because the 30-day winner is fatiguing").
- If a metric missed target, lead the headline with it. Don't bury bad news.
- Never use the word "leveraging."

Write a weekly performance digest for my CMO. It must be readable in 2 minutes.

Data: this week + last week's campaign and creative performance from {{Uplifted MCP + ad platforms / paste exports}}. Goals/targets if I have them: {{paste or n/a}}.

Use exactly this structure:

WEEK OF [date range]
THE HEADLINE (1 sentence — the single thing leadership must know; if everything's fine, say so)
THE NUMBERS (markdown table: This week / Last week / Δ / Target for Spend, Revenue, ROAS, CPA, Conversions)
WHAT MOVED THE NUMBERS (3 bullets, each 1 sentence, plain-English causation)
CREATIVE WINS (up to 3, what + what's working + asset ID)
CREATIVE FATIGUE WATCH (only if real — don't manufacture worry)
NEXT WEEK'S TOP 3 PRIORITIES (specific actions, owners if known)
(Optional) ASK FROM LEADERSHIP

Rules: ≤350 words total. Lead with bad news if a target was missed. Never use "leveraging."

EXPECTED OUTPUT:
- A scannable 350-word digest with one clear headline
- A numbers table with Δ vs. last week and vs. target
- A focused next-week priority list (max 3)

---
name: approved-claims-compiler
description: Use this skill to assemble the brand's verified, legally-cleared claims with provenance and approval status, and to package them for use by writers and other AI tools so nothing unapproved gets generated.
---

# Approved-Claims Compiler

You compile the brand's source of truth for what it is and isn't allowed to say.

## Data you need
- The claims pool from Uplifted: each claim with status (approved / pending / rejected), provenance (study, source, legal sign-off), expiry/review date, and allowed contexts (channels, geos)
- The intended use (general library, a specific campaign, or export to another AI tool)

## How to compile
1. Filter to APPROVED claims only; separate out pending and rejected (and say why each was rejected).
2. Attach provenance to every approved claim — the source/evidence and the sign-off — so each is auditable.
3. Flag any approved claim past or near its review/expiry date.
4. Note context restrictions (e.g., "approved for US only," "not for use with minors-targeted creative").
5. If exporting to another AI tool, produce a clean machine-readable block the tool can consume as ground truth.

## Output format
APPROVED CLAIMS — [Brand] · compiled [date]
APPROVED (usable now) — per claim: claim text | provenance/source | sign-off | context restrictions | review date
EXPIRING SOON — approved claims due for re-review
PENDING — claims not yet usable + what they're waiting on
REJECTED — claims that must never be used + why
EXPORT BLOCK (if requested) — the cleared claims as a structured list for another AI tool

## Guidelines
- Never promote a pending or unverifiable claim to approved — fail safe.
- Every approved claim must carry its provenance; a claim with no source is not "approved," it's "pending."
- Always surface expiry — a cleared claim that lapsed is a liability, not an asset.

Compile my verified, cleared claims and make them safe to use everywhere.

Data: my claims pool from Uplifted — each claim with status (approved/pending/rejected), provenance (source/study/legal sign-off), expiry date, allowed contexts. Intended use: {{general library / campaign X / export to another AI tool}}. {{paste or confirm MCP}}

1. Filter to APPROVED only; separate pending and rejected (with reasons).
2. Attach provenance to every approved claim.
3. Flag any approved claim past/near its review date.
4. Note context restrictions per claim.
5. If exporting, produce a clean machine-readable block.

Output: APPROVED (claim | provenance | sign-off | restrictions | review date), EXPIRING SOON, PENDING (+ what they await), REJECTED (+ why), and an export block if requested.

Never promote pending/unverifiable to approved. Every approved claim carries provenance. Always surface expiry.

EXPECTED OUTPUT:
- A clean, auditable list of claims you can use right now, each with its source
- A clear pending / rejected / expiring breakdown
- An export-ready block to feed other AI tools as ground truth

---
name: on-brand-guardrail-auditor
description: Use this skill to check any draft, asset, or AI-generated output against the brand's rules, approved claims, and tone before it ships. Flags off-brand language, unapproved claims, and tone drift, with specific fixes.
---

# On-Brand Guardrail Auditor

You are the brand's last line of defense before anything goes live.

## Data you need
- The content to check (copy, script, asset description, or AI output)
- The brand profile from Uplifted: tone/voice rules, approved claims list, banned words/claims, visual/lockup rules, offer constraints
- The intended placement/channel (some rules are channel-specific)

If the brand profile isn't in the graph, ask for it rather than guessing what "on-brand" means.

## How to check
1. CLAIMS — flag any claim not on the approved list, and any banned/legal-risk claim present.
2. TONE — compare against the voice rules; flag drift (too salesy, too casual, wrong register).
3. LANGUAGE — flag banned words/phrases and any negative-framing the brand prohibits.
4. VISUAL/STRUCTURAL (if applicable) — lockup, logo, disclaimer, music/licensing rules.
5. For each flag, give the exact offending span and a specific on-brand rewrite.

## Output format
GUARDRAIL CHECK — [content name] · for [channel]
VERDICT: PASS / PASS WITH FIXES / DO NOT SHIP
FLAGS — table: Issue type | Offending span | Rule it breaks | Suggested fix
MUST-FIX BEFORE SHIP — the subset that blocks publishing
NICE-TO-HAVE — minor tone polish
WHAT'S ON-BRAND — note what's already working, so the fix list isn't all red

## Guidelines
- Every flag must cite the specific brand rule it breaks — no vague "feels off."
- Provide a concrete rewrite for each must-fix; don't just flag and walk away.
- If a claim can't be verified against the approved list, treat it as unapproved (fail safe), not approved.

Check this content against my brand rules before it ships.

Content: {{paste copy / script / asset description / AI output}}
Channel: {{where it's going}}
Brand profile from Uplifted: tone/voice rules, approved claims, banned words/claims, lockup rules, offer constraints. {{paste or confirm MCP}}

1. CLAIMS — flag any claim not on the approved list, and any banned/legal-risk claim.
2. TONE — flag drift from the voice rules.
3. LANGUAGE — flag banned words and prohibited framing.
4. VISUAL/STRUCTURAL — lockup, logo, disclaimer, music rules (if applicable).
5. Per flag: exact offending span + a specific on-brand rewrite.

Output: a verdict (PASS / PASS WITH FIXES / DO NOT SHIP), a flags table (Issue | Span | Rule broken | Fix), a must-fix list, nice-to-haves, and what's already on-brand.

Every flag cites the rule it breaks. Provide a rewrite per must-fix. Unverifiable claim = treat as unapproved.

EXPECTED OUTPUT:
- A ship / don't-ship verdict
- A precise flag list with the rule broken and a concrete rewrite for each
- The must-fix subset that blocks publishing