Overview
What it does
Breaks a winning ad into timecoded scenes, attributes the result to the moment that carried it, and hands back the raw clips to rebuild from.
When to use it
You have a winner and want to know which exact moment carried it — the hook, the demo, the proof, the CTA, the 3-second product reveal — so you reuse the segment that worked, not just the whole ad. Also use it to pull the raw clips behind a winner so an editor can rebuild from source.
Pairs with MCP
Uplifted attaches performance at the scene level and keeps every raw clip linked to the final it produced. No other tool can tell you which 3 seconds moved the number, or hand you the source clip to rebuild from — this is the moment-level moat.
Best for
Creative strategists, video editors, performance marketers.
Install
Drop this entire block into a SKILL.md file inside your Claude project’s .claude/skills/scene-level-teardown/ folder. Claude auto-invokes it when you ask which part of an ad worked, or for the raw clips behind a winner.
---
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.Prompt
Breaks a winning video into timecoded scenes, attributes the result to the exact moment that carried it, and hands back the clips to rebuild from. Use it to reverse-engineer why an ad worked instead of guessing.
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