| name | net-new-video-editor |
| description | Turn newly recorded talking-head footage into review-ready vertical video drafts with an explicit edit plan, deterministic FFmpeg rendering, captions, hook cards, audio normalization, and visual QA. Use for original Instagram Reels, TikTok videos, YouTube Shorts, LinkedIn videos, founder-led recordings, multiple takes of a new script or idea card, or requests to automate the first video-editing pass. Do not use to mine clips from long-form source videos. |
Net-New Video Editor
Create a reversible first edit from fresh recordings. Keep creative decisions in JSON and pixel operations in the bundled renderer.
Preamble
Run from the repository root when the optional shared telemetry helpers are present:
python3 telemetry/version_check.py 2>/dev/null || true
python3 telemetry/telemetry_init.py 2>/dev/null || true
Remote telemetry is opt-in and never includes content, file paths, repository names, or credentials.
Establish the package
Read references/project-contract.md. Locate the exact source recordings, transcript, idea card or brief, proof assets, and screen recordings. Never substitute another recording, brand, account, or asset library.
Initialize a new project only when the destination is clear:
python3 scripts/net_new_video_editor.py init --project <project-dir>
Copy or point only user-authorized inputs into the generated package. Preserve originals.
Inspect before editing
Run:
python3 scripts/net_new_video_editor.py inspect --project <project-dir>
Review intake-report.json. Stop when the package has no playable take, the requested target does not match the supplied footage, or required external assets are missing.
Build the edit plan
Use the transcript and brief to create edit-plan-clean.json. Treat the spoken hook and claim boundaries as ground truth.
- Select one source take explicitly.
- Keep segment order intentional and timestamps within the source duration.
- Remove clear false starts, long dead space, and isolated filler only when the cut remains natural.
- Preserve breaths that help meaning.
- Put the hook on screen for at most five seconds.
- Use captions in short readable phrases.
- Add proof or screen inserts only when supplied and relevant. The bundled renderer handles the base assembly; add complex overlays in a separate, documented pass.
- Normalize speech without clipping.
For a second version, copy the plan to edit-plan-aggressive.json and make only named retention edits. Do not silently change factual claims.
Validate each plan:
python3 scripts/net_new_video_editor.py validate --project <project-dir> --plan <plan.json>
Render deterministically
Run the renderer from this skill directory:
python3 scripts/net_new_video_editor.py render \
--project <project-dir> \
--plan <plan.json>
The renderer trims paired audio and video, concatenates the selected segments, creates a center-safe 9:16 frame, rasterizes captions and the hook card with Pillow, composites them with FFmpeg, normalizes audio, writes H.264/AAC MP4, and creates three QA frames plus qa-report.json.
Use --dry-run to inspect the FFmpeg command. Use --force only when replacing the exact derived export is intended.
Review the output
Inspect the exported MP4, all three QA frames, qa-report.json, and the plan diff between variants. Verify:
- audio and mouth movement stay synchronized after every cut;
- no word starts or ends abruptly;
- captions match the speech and stay inside safe margins;
- the crop keeps the speaker visible;
- the hook is legible and repaid by the video;
- loudness is consistent and peaks do not distort;
- every proof insert and numerical claim has a supplied source;
- the export is 1080x1920 H.264/AAC unless the brief requires another format.
Return the plan, export paths, QA evidence, and any required pickups. Never publish, upload, delete originals, or overwrite an approved master without current explicit approval.
Completion states
DONE: both render and QA pass, and the review artifacts exist.DONE_WITH_CONCERNS: the draft is usable but a named creative or source concern remains.NEEDS_CONTEXT: a take, transcript, brief, or approved asset is missing.BLOCKED: FFmpeg, source access, or format validation prevents a safe render.
