Ad video editor
Read video-edit-rules first. Its rules apply to every step here.
The goal: ads that stop the scroll in the first second and say only what the speaker said. The test: every export passes the quality gate and the claims table, and each hook is its own file on the same body.
Organic reels go to talking-head-reel or reel-editor. This skill adds what ads need on top.
Step 1. The edit plan
Write work/ad-plan.md before you touch the footage.
- The body: the lines, in order, with source timestamps.
- The hooks: 2 to 5 opening lines to test. Each one is either a separate hook take, or a line from the body moved to the front (a cold open).
- The versions to export: sizes and hooks.
- The claims table (step 2).
- Anything you are unsure of: a word you cannot hear, a number that seems off, a line that may not be approved. Flag it. Do not fix it silently.
Never drop, reorder or reword an approved script line without the user's yes.
Step 2. Claims stay inside the words
Ads carry legal and platform risk. Treat every on-screen word as a claim.
| On screen | Spoken by | Source time | Same meaning? |
|---|---|---|---|
| "Lost 12 Kg In 4 Months" | speaker | 00:14.2 | yes, says "I lost twelve kilos in four months" |
Rules:
- Every on-screen word, number, price and name maps to a spoken line in this footage.
- A label may shorten a line. It may never add to it, round a number up, or make it stronger.
- A figure given as an example gets a small "Example" tag.
- Results claims (health, money, income) follow the platform's ad policy. Flag them for the user to check.
- A line with no source stops the render.
Step 3. The body
- Cut the body once, following the cut rules: silence by level, real pauses between thoughts.
- Pace: ads run tighter than organic posts. Cut pauses over about 0.2 s inside a thought. Keep 0.25 s between thoughts.
- Hide each join under a framing change (1.0x, 1.15x, 1.32x) or a scene cut.
- Captions on every spoken word after the hook. Viewers watch with the sound off.
- Scenes may take more of the runtime than in organic posts (often 30 to 45%). Say the share in the report.
- End on the spoken call to action and its last word, plus 0.3 to 0.5 s. No silent tail.
Step 4. Hooks
- The hook text is on screen on frame 0 or by 0.15 s at the latest.
- The first frame reads without sound: big text plus a face, or a clear picture of the problem.
- By 3 s the viewer knows the problem or the product.
- Each hook is its own export with its own hook end time. The body after it stays the same frames.
- Cold-open hooks use only the speaker's words from the body.
- Name the files so the hook and the size are clear:
out/<ad>_h1_9x16-v01.mp4,out/<ad>_h2_4x5-v01.mp4.
Step 5. Sizes
Build the 9:16 master first. Then reframe.
| Size | Pixels | Safe text area | Notes |
|---|---|---|---|
| 9:16 | 1080x1920 | y 220 to 1470; right 140 px clear below y 1100 | Stories, Reels, TikTok |
| 4:5 | 1080x1350 | about 60 px margin each side, 120 px at the bottom | Feed |
| 1:1 | 1080x1080 | about 60 px margin each side | Some feed placements |
For each new size:
- Re-place the crop so the face keeps its share of the frame.
- Re-place captions under the chin, inside that size's safe area.
- Scale scenes from the HTML, do not crop the 9:16 render.
- Run the layout check on that size.
- Look at one frame per scene and one caption frame in each size before the full render.
Check the platform's current ad specs before the first export for a new account. They change.
Step 6. Sound
- Voice at -14 LUFS, true peak at or under -1.5 dBTP.
- Soft pops only, at least 20 dB under the voice. No whoosh or riser.
- Optional bed 20 LU under the voice, fading on a beat.
Step 7. Quality gate before review
Nobody reviews an ad that has not passed this list. Run it on each exact export.
- Frame scan:
python $KIT/scripts/frame_scan.py <file>. - Layout check on every scene, in every size.
- Word check: re-transcribe the export and compare it with the plan, word for word.
- Edge check on every join.
- Claims table: every on-screen word has a source line.
- Loudness:
python $KIT/scripts/loudness.py <file> --target -14. - Sound list with times and levels.
- Hook check: the first frame, the 0.15 s frame and the 3 s frame, saved as stills.
A fresh reviewer (a new agent with no build context) can score the export out of 10 against this list. Fix and re-run until it scores 8 or more, or until three rounds pass. Then show the user.
Footage problems
- A wrong word, a stumble that cannot be cut, or a bad take goes back to whoever made the footage. Do not hide it under b-roll.
- If the footage is AI-made, label it as the platform requires.
Report
- Each export: file, size, hook, length, first line.
- Claims table.
- Scene share of runtime.
- Every sound effect with time and level.
- Gate results per file, and anything not checked.