agentleFS
kajisho5kajisho5/video-editing-skill/SKILL.md

video-editing-skill

Skill12 starsChanged 6 days agoMITView on GitHub

Deterministic video editing from a typed edit request (trim, cut, concat with transitions, speed change, fit / fill / resize, still-image overlay) with an exact source-to-timeline mapping and per-operation provenance. Executes through ffmpeg-skill; never takes commands, argv, filter strings or executables. Use when an agent has already decided what to edit and needs the edit performed, verified and traced. Not for deciding what to cut, transcribing, captioning, colour grading or audio mastering.

2 other projects use a version of this.

Adopted by

Most recently taken up by

  • Other versions2 projects use a version of this.

    Word for word · 1

    1. trku052-png/social-media-youtube-skills★ 1100% same

    Edited · 0

    None found.

    Wrote their own version · 1

    1. social-media-skills/video-editing-skill★ 4same name

    How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.

  • In this repositoryOne other file it gives its agents.

The file

Files
SKILL.md

video-editing-skill

video-editing run - --json --workspace DIR [--allowed-input ROOT] reads one request document on stdin and prints one response document on stdout. Everything else (skill, doctor, validate, plan) is documented in README.md; the machine-readable contract is video-editing contract --json.

Workflow for a caller:

  1. doctor --json --workspace DIR once: ok must be true; operations[] says per type whether it is AVAILABLE (tool and the flags it needs present in the installed ffmpeg-skill, any release; encoders / filters found by its doctor) and supported_operations lists exactly those.
  2. Build the request from the contract's request_shape: sources (files under an allowed root), allowlisted operations with typed params, outputs (relative paths under the workspace).
  3. plan - --json to see the operation graph, the timeline mapping, the tool per step and the commands ffmpeg-skill would run. Nothing is written.
  4. run - --json. On success every output carries its sha256, size, timeline and an OBSERVED probe; every source its probe; every operation a provenance record (completed / reused); the document is self-checked against the contract shape before it is printed. On failure {"ok": false, "error": {code, message, retryable, details}}, the failed record with its error, later operations skipped, and no output file is left behind.

Overlay boundary (ADR-035): OVERLAY is plain still compositing as part of an edit (one png / jpg at a named position for a time range, alpha respected). A video layer, picture-in-picture, chroma key, a logo or image treated as a graphic (fade, theme, safe zones) and any moving graphic belong to motion-graphics-skill (image_overlay, logo, video_overlay, rendered_layer); ffmpeg-skill is the engine under both. A video given as OVERLAY.params.image (a video source or an operation) is refused before anything runs as INVALID_INPUT overlay_video_layer, naming motion-graphics-skill video_overlay (ADR-036).

Media rules (contract.media_compatibility): every source is probed first; OVERLAY needs a video input with an audio stream; CONCAT conforms sizes / rates and adds audio when any input has it. Engine gaps (a missing ffmpeg-skill tool, encoder or filter) are refused before execution as TOOL_ERROR.

Operation types (the allowlist): TRIM, CUT, CONCAT (with params.transition: an xfade type, or {type: stinger, stinger, cut_point, cover} for a stinger with alpha, the switch under full cover at its cut point; read cut_point / cover from its stinger.json; the cover at the cut point is measured and a stinger that does not cover it is refused), SPEED (factor, optional smooth: blend | interpolate for slow-motion), FIT, FILL, RESIZE, OVERLAY, ROTATE (turn 90/180/270 clockwise and/or mirror h/v), FREEZE (hold seconds of one frame: at at, or the last frame when at is omitted), REVERSE (backwards, picture and audio), IMAGE_INSERT (one kind: "image" source → a silent clip of duration seconds, optional width / height / fps, that CONCAT can join: a title card or end slate), CROP (exactly the pixel rectangle {x, y, width, height} you give: even sizes, inside the frame as displayed; decide the rectangle yourself, e.g. from ffmpeg-skill cropdetect), CONDENSE (segments [{start, end, label}] and rules {label: cut | keep | {speed: f}}, default keep: apply the proposals from video-editing propose that a person approved; e.g. cut retakes and keep silences at double speed), PUNCH_IN (start, end, scale in (1, 4], subject {x, y, width, height}: a hard-cut zoom on the box you give, e.g. detection-skill's face box; which jump cuts get one is your decision). FREEZE / REVERSE / IMAGE_INSERT / CROP need ffmpeg-skill's freeze.py / reverse.py / insert.py / crop.py; doctor --json says whether the installed checkout has them. Anything else (POSITION included) is refused with UNSUPPORTED_OPERATION; the contract's unsupported list says why.

TRIM / CUT take precision: frame (default: re-encoded, frame-accurate), keyframe (lossless, may land on a keyframe), auto (a stream copy where it lands within 0.2 s, else re-encoded) or smart (frame-accurate, only the partial GOPs at each cut point re-encoded; for long recordings cut between keyframes). smart needs ffmpeg-skill cut.py --smart (3.2.3+); without it, or for a source the engine cannot match, the whole cut is re-encoded and execution.operations[].cuts.smart says why (applied: false, reason_code).

Times are exact: "1:30", "00:01:30.250", {"frames": 300, "fps": "30000/1001"} or a number of seconds.

Frames are exact too: RESIZE keeps the aspect (width → height = even(width × sh / sw)), FIT pads to an aspect, FILL crops to it (optionally off-centre via anchor: {x, y}, each 0..1); ROTATE keeps every pixel losslessly (degrees: 90 | 180 | 270, flip: h | v), swapping width and height only for a 90/270 turn; the target frame is reported before execution (normalized.target_frame) and verified after. outputs[].encoding may set crf (14..28) and preset (x264 vocabulary); nothing else about encoding is configurable (contract.encoding).

contract --check tests/contract/contract.json verifies the live contract against the implementation, the docs and that saved copy; exit 1 on any problem.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.