Make feedback locatable
Video Recap separates new creation, execution of a direction and revision of an existing version. Latest feedback governs revision; unspecified elements stay fixed. It does not require new story hypotheses every round.
Original example: viewers cannot understand a letter handoff. Instead of “too fast,” say: “Before the handoff at 00:21, restore the original line naming the proxy. Keep other narration, music and the ending.” Use the current candidate’s actual times in production.
When a premise is missing, recover source evidence or add necessary supported context. Do not invent an earlier agreement merely to smooth the edit.
Record changes and retained decisions together
| Layer | Record |
|---|---|
| Story | Change: Restore the handoff reason Retain: Recipient’s recognition order |
| Picture | Change: Restore a specified source interval Retain: Other shot order and reactions |
| Narration | Change: Revise only the affected sentence if necessary Retain: Other text and voice |
| Sound | Change: Make room for restored dialogue Retain: Adopted music and other source passages |
| Subtitles | Change: Update affected cues and positions Retain: Other wording and styling |
Dependencies still matter. Added dialogue may move later positions. Retaining decisions means identifying required movement rather than using it as an excuse to rewrite unrelated content.
Update the decision record that owns the change
The skill records expression, rhythm and subtitle feedback in the style record, and shot/sound feedback in the audiovisual record. Update the story plan only for a changed storyline. Remove obsolete plan descriptions when deleting material so later runs do not restore it.
A useful instruction: “Restore only this dialogue and its necessary duration. Identify affected positions and explain unavoidable movement; retain other wording and sound decisions.” The production workflow owns the files; authors need not use internal fields as creative subjects.
Explicit latest decisions supersede old ones. For vague feedback, propose a concrete local repair on the existing version instead of automatically restyling the whole film.
An Eight-Second Repair That Keeps Later Positions
This is a paper exercise without a rendered or listened-to result. Output 16–20 uses source 30–34 of walking; source 33.2–34 is repetitive. Output 20–24 uses source 40–44 for a letter handover at output 21. Source 39.2–40 already contains the complete “Give it to Secretary Zhou” at a natural boundary, omitted by the old in-point.
| Clip | Old plan and candidate |
|---|---|
| Walking | Source 30–34 → output 16–20; candidate source 30–33.2 → 16–19.2, retaining necessary approach and space. |
| Letter | Source 40–44 → output 20–24; candidate source 39.2–44 → 19.2–24, naming first, then the original handover and recipient reaction. |
| After 24 | Positions remain; retain unchanged narration text and WAVs, checking affected captions only. |
3.2 plus 4.8 remains eight seconds, so this proposal need not shift the later video. Without genuinely disposable walking, do not copy the numbers. Extend with declared dependencies or select another existing complete premise. Do not remove necessary approach just to fit.
In the same cut plan, protect source 39.2–40 as audio evidence and the needed 40–44 handover/reception as video evidence, keeping premise before reply. Enter only moments actually present. Coverage and ordering checks do not prove audible clarity.
Restored dialogue changes sound at 19.2–20. Retain chosen music, its original start, and other sound decisions, then listen for masking. Report any necessary local music-gain change. This is not an unchanged encoded full mix. A requirement to freeze the whole mix is incompatible with adding dialogue and needs a different revision target.
After producing the candidate, hear the name, watch the handover and reply, then compare unaffected content before 16 and after 24. Equal length establishes a clock plan rather than complete source, audio, or caption verification.
Verify both the change and retained material
Play the changed passage normally: the premise is audible and the handoff makes sense. Review adjacent joins for lost reactions and subtitle drift. Watch the complete candidate and listen separately for repeated mix processing.
Compare the retained list with the old version: unmentioned text, voice, music position and level, and the ending’s question. Screenshots and logs aid location but do not replace comparison of actual media.
Report actual changes and necessary dependencies. Unrun or unheard items remain unverified; identical parameters do not prove identical overall effect.
FAQ
Must a subtitle change rewrite narration?
Usually not. Identify wording, timing or appearance and update that layer and required dependencies. If meaning changes, explicitly revise narration and audio too.
Do it with the skill
Ask video-script to revise the current version with a change/retain list, locatable repairs and separate checks for changed and unchanged passages.
Read the method: Local edit revision and decision records