skills/creativeclaw/references/workflows/edit-video.md

# Edit existing footage

Preserve the original and create clearly named derivatives. Distinguish deterministic editing, generative source-video editing, and continuation before choosing a model. Do not regenerate untouched footage merely because a generative model is involved.

## Route the operation

| Need                                               | Tool or model path                                 |
| -------------------------------------------------- | -------------------------------------------------- |
| Cut a time range                                   | `trim_video`                                       |
| Resize, crop, or pad                               | `scale_video`                                      |
| Burn captions                                      | `add_subtitles`                                    |
| Extract first, middle, last, or regularly spaced frames | `extract_frames`                             |
| Concatenate clips or audio; add audio to video     | `merge_media`                                      |
| Remove a background                                | `remove_background`                                |
| Upscale                                            | `upscale_media`                                    |
| Transcribe with timings                            | `transcribe`                                       |
| Clean a voice track                                | `isolate_audio`                                    |
| Generative source-video edit                       | `generate_video` with a compatible source-edit mode |
| Add new action before or after existing footage    | Generate only the continuation, then `merge_media`  |

## Generative edit routing

Call `list_models({ category: "video" })` and `get_model_params` before choosing or submitting a generative edit. Runtime capability and limits are authoritative.

- For a source clip up to 10 seconds, default to `video/gemini-omni-flash`. It is the best-value Creative Claw route for one targeted natural-language change while retaining the source timeline. Use one clear change followed by an explicit preservation sentence.
- For a 4–30 second source edit, a long continuation, or a reference-rich transformation, use `video/seedance-2.5`. Set `extras.omni_reference_task_type` explicitly to `edit` or `extend`; use `aspect_ratio: "auto"` and the duration contract returned by `get_model_params`.
- Use `video/seedance-2.0-mini`, presented as **Seedance Mini**, only for an explicitly cost-sensitive draft whose source and output fit its current limits. It reconstructs from references rather than guaranteeing a surgical edit, so do not use it when exact frame, logo, uniform, text, or identity preservation is critical.
- Use `video/flux-3` only for continuation when its current schema explicitly exposes source-video extension. Do not describe it as a source editor unless `get_model_params` exposes an edit mode.

Never recommend or proactively route to an LTX or DreamActor model. If the user explicitly requests one, preserve their choice but do not present it as a recommended route.

## Preservation-first continuation

When the user says to leave the original video unchanged and add something before or after it:

1. Keep the complete original video untouched.
2. Trim only the smallest useful head or tail context accepted by the continuation model.
3. Generate one continuation from that context.
4. Concatenate the untouched original and the approved continuation with `merge_media`.
5. Preserve the original audio on the original span. Add or mix audio only for the new span unless the user requested a broader audio change.

This is a continuation job, not a full-source edit. Do not spend credits regenerating the original timeline. For an original clip longer than the model's source limit, trimming context is expected and does not require splitting the untouched portion into generative segments.

## Long edits and exact details

- If only one interval needs a generative change, trim that interval, edit it, and merge it back between untouched source spans. Do not process the entire source through the model.
- If multiple separated intervals need changes, propose the exact intervals and number of paid generations before submitting them.
- A model rejection, duration-limit failure, insufficient balance, or estimate does not authorize switching models or creating another take. Report the constraint and the best compatible route.
- Treat exact logos, jersey numbers, readable text, scoreboards, and brand marks as deterministic compositing or VFX requirements when they must remain exact. Generative preservation language reduces drift but cannot guarantee pixel-accurate details.

## Transcript-driven cut

1. Import the source through `../platform-upload.md`.
2. Use `transcribe` and resolve the job when word timings are required.
3. Select cut points on word or sentence boundaries.
4. Use `trim_video` for approved ranges.
5. Scale/reframe, then add subtitles only after timing is locked.
6. Merge approved clips and audio.

## Audio polish

Use `isolate_audio` before transcription or final assembly when the recording contains material noise, music, or reverb. Keep the untouched source. Resolve the queued result before passing it downstream.

## Quality gate

Check cut boundaries, audio sync, subtitle timing and safe zones, crop on faces/products, resolution, duration, and continuity between merged clips. Give every derivative a name and tags that distinguish it from its source.