Raibit-AI/ad-teardown

Ad teardown

1. Measure (free, local)

Run the bundled script on each video. It needs ffmpeg; whisper.cpp (whisper-cli) is optional, for transcripts.

bash scripts/teardown.sh path/to/ad.mp4 out_dir/

It writes:

• hook.jpg: frames every 0.5 s from 0 to 3 s.
• sheet.jpg: 16 frames across the whole ad.
• metrics.json: duration, cuts, cuts_per_min, avg_shot, first_cut.
• transcript.txt: only if whisper-cli is installed. Set WHISPER_MODEL to the model's path.

2. Look and tag

Read hook.jpg and sheet.jpg. Ignore the yellow timestamps; they are not part of the ad. Fill in this record, judging only what you can see or read:

{
 "format": "food_montage | owner_or_staff_talking | creator_or_influencer | customer_reactions | offer_graphic | kitchen_process | mixed",
 "frame0": "what is literally in the first frame",
 "hook_type": "food_motion | food_still | face_talking | text_headline | offer | person_eating | process_cooking | other",
 "hook_text": "on-screen words in the first 3 s, or null",
 "motion_in_first_second": true,
 "hands_in_frame": "none | some | most",
 "people": "none | staff_or_chef | owner | customers | creator",
 "speech": "none | voiceover | on_camera_talking",
 "onscreen_text": "none | dish_labels | headline | offer | captions | mixed",
 "offer": "the deal in words, or null",
 "brand_shown_by_sec": 0.0,
 "ending": "logo_card | offer_card | food_shot | person_talking | abrupt | other",
 "cta_words": "exact call to action, or null",
 "visual_style": "phone_native_raw | polished_studio | mixed"
}

3. Compare with the benchmarks

Open benchmarks.md and report the ad against it. Pass or miss on each line, with the gap in numbers. For example: "first cut at 4.1 s against a 1.9 s benchmark", or "no hand in frame, while 79% of winners have one".

4. Report

Keep it short and plain:

1. Verdict in one line.
2. What it does that winners do. Up to 3 points.
3. The 3 biggest gaps, each tied to a timestamp and a concrete fix.
4. One thing worth stealing from it, as a pattern rather than footage.

When the user gives several ads, add a summary table and the patterns shared by 3 or more of them.