Communitygithub.com

cut-as-code

>

cut-as-code とは?

cut-as-code is a Claude Code agent skill that >.

対応Claude Code~Codex CLI~Cursor
npx skills add https://github.com/WhiteTowerAI/cut-as-code/tree/main/skills/cut-as-code

お気に入りのAIに質問する

このエージェントスキルを事前に読み込んだ状態で新しいチャットを開きます。

ドキュメント

Cut as Code Entry Point

Use this as the mandatory entry point for cut-as-code video work. Own routing and project-state recovery only. The selected leaf skill owns editorial decisions, review gates, scripts, rendering details, and self-checks.

1. Start From Project State

Apply the first matching row. Do not evaluate lower rows.

StateAction
The user names a specific operation on an existing projectInspect only the state needed for that operation, then load its owning skill. Skip general intake.
An explicit project root contains work/project.jsonRead work/project.json, work/timeline.json, and only the selected operation's plan/evidence. Resume from recorded state. If the owner is not unique, ask one routing-only question.
The user supplies a source file or URL, no project, and a clear deliverableLoad /video-understand to initialize the project and shared evidence, then continue to the requested operation.
No project/source or no intended deliverable is identifiableAsk one concise question requesting the missing project/source and desired result. Do not start a skill yet.

The project root must be explicit; never treat the repository root as a video project. Treat work/project.json as the only shared manifest. Determine progress from operation status, revision, based_on, check, and render state, not from directory presence. Do not re-run verified work. If a dependency is missing or stale, complete that dependency first and then resume the requested operation.

Routing is read-only. Never repair, reset, approve, remove, or activate an operation while deciding where to send the task.

2. Route by Deliverable

Match the requested output, not a keyword mentioned in passing.

Requested resultOwning skill or workflow
Probe, transcript, objective analysis, semantic understanding, or new project initialization/video-understand
Tighten, shorten, remove weak sections, or create a first/rough cut/video-cut after /video-understand
Correct color, create/select a look, or deliver/apply a LUT/video-color-grade; use its standalone path only when the user explicitly wants no Project Protocol delivery
Add transcript-timed visual cutaways/video-add-b-roll after /video-understand
Add sourced motion graphics or animated semantic overlays/video-add-graphic-motion after /video-understand
Add titles, lower-thirds, statistics, quotes, chapters, or calls to action/video-add-content-cards after /video-understand
Add subtitles, captions, or karaoke captions/video-add-captions after /video-understand
Compare the original with actual final pixels in source time/video-edit-compare after the verified main delivery exists
Select, review, and plan short-form ranges/video-to-shorts after /video-understand; do this before requested content cards and captions
Finalize and extract approved horizontal or vertical short-form derivativesResume /video-to-shorts after the verified main delivery exists
Compile or render an existing Project Protocol deliveryUse build_render_plan.py and render_project.py from /video-understand after every selected active operation passes its own gate

A generic request such as "edit this video" does not uniquely select /video-cut. Ask one routing question unless the desired output is clear from existing project state or the surrounding request.

3. Orchestrate Multi-Operation Requests

Freeze only the operations the user requested; never turn the route table into a default package. Run the shared understanding prerequisite once. Execute selected main-sequence operations in canonical order:

cut -> color-grade -> b-roll -> captions -> content-cards -> graphic-motion

For a request that includes Shorts, content cards, or captions, use this staged flow:

/video-understand
-> /video-to-shorts plan and candidate review only
-> /video-add-captions
-> /video-add-content-cards
-> /video-add-graphic-motion
-> compile and render the shared main delivery once
-> /video-to-shorts finalize and extract approved derivatives

Complete any requested timeline-changing operation, such as /video-cut, before Shorts planning; a changed work/timeline.json invalidates candidate approval. Shorts planning writes shared plan/review state and must not create final.mp4 or anything under final/shorts/. Captions, content cards, and graphic motion then add their own shared plans and overlay contributions in that relative order. The same order applies when only any pair is selected: captions first, content cards second, graphic motion last. Re-read work/project.json before each handoff. Do not run leaf skills concurrently when they can write work/project.json, work/timeline.json, an operation plan, or shared render artifacts. Build and render the main delivery once, after all selected active operations validate. Run /video-edit-compare and only the finalization/extraction phase of /video-to-shorts against that verified delivery; shorts remain outside sequences.main.operations.

For /video-add-content-cards and /video-add-graphic-motion, reserve the visible face and head silhouette of every primary or foreground person, speaker, presenter, interviewee, or semantically important person for the complete cue. Incidental background-only people who are not a narrative or visual focus are exempt; protect uncertain cases. If an overlay touches a protected face or head, reposition it first, then scale or redesign it; skip the cue when no compliant placement exists.

An all-skipped or zero-cue result is valid when the owning skill permits it. Do not invent filler work to keep an operation active.

4. Hand Off, Then Leave

Read the selected skill's complete SKILL.md and every required sub-skill before acting. For one operation, hand off once and leave this entry point. For a requested multi-operation run, enter one owner at a time; after its self-check, recover project state and continue to the next frozen operation without reopening intake.

Do not ask the user to approve a route that is already unambiguous. User interaction belongs only to missing inputs, genuine routing ambiguity, credentials, or the editorial/review decisions required by the owning skill. Never substitute an Agent decision for a required human decision unless the user explicitly delegates it.

Example

For "continue this understood and cut project by adding captions," read the current project and timeline revisions, then enter /video-add-captions. Do not re-transcribe, re-cut, or ask whether captions are the intended route. An unrelated pending operation does not change the route, though an active unverified operation may block the final shared render and must be reported by the compiler.

Failure Rules

  • Invalid project state: report the exact validation problem; never reinitialize over it.
  • Missing leaf skill or prerequisite: stop and name it; do not reconstruct it from memory.
  • Missing source/project path: ask for it in the single routing question.
  • Conflicting requested outputs: ask which deliverable owns the current run.
  • Existing user changes or artifacts: preserve them and work with the recorded state.

Individual skills in this repo

This repo contains 8 individual skills — each has its own dedicated page.

関連スキル