CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Git

🔒 This section must never be modified. Leave it byte-for-byte unchanged.

• Prefer small PRs focused on one feature, one bug fix, one chore, or one refactor.
• Use branches:
  • feat/<short-name>
  • fix/<short-name>
  • chore/<short-name>
  • refactor/<short-name>
  • docs/<short-name>
• Use Conventional Commit style for PR titles:
  • feat(editor): add marquee selection
  • fix(runtime): handle missing tween state
  • docs(agents): update collaboration rules
• PR descriptions must include:
  • What changed
  • Why it changed
  • How it was checked
• Default to squash merge.

What this repo is

cut-as-code is a stack of agentic video-editing skills plus its local Editor Plugin and browser UI. Each directory under skills/<name>/ is a self-contained skill: a SKILL.md (the agent playbook — read it first), plus scripts/, examples/, and reference/. Plugin runtime code lives under .codex-plugin/, runtime/, and ui/; the stable desktop bootstrap lives under launcher/. There is no root build, package manifest, lint config, or aggregate test suite. Scripts and checks run ad hoc. When you change a skill, the SKILL.md is the spec; keep it and its scripts in sync.

Each SKILL.md starts with YAML frontmatter (name:, description:) — the name is the slash-command trigger for that skill.

Skills

Shared project protocol V1

• Skills are optional and composable; there is no fixed global pipeline. A project can run cards directly or cut → color grade → cards.
• work/project.json is the only shared manifest. It records operation dependencies, statuses, render contributions, integer revision values, and based_on checks.
• work/timeline.json is the custom one-source, chronological source-to-program mapping. V1 does not use OpenTimelineIO or support reordered/duplicated clips or nonlinear speed.
• Canonical time values are seconds. All ranges are half-open [start_s, end_s) and use explicit source_range and program_range objects.
• Each operation declares a target (sequence and scope) and effects such as timeline, pixel, geometry, audio, or added-track changes.
• Render contribution kinds are timeline-transform, video-filter, audio-filter, overlay, precomputed-asset, and output-constraint.
• Domain decisions remain in work/cut/edit-plan.json, work/color-grade/grade-plan.json, work/b-roll/broll-plan.json, work/motion-graphics/motion-graphics-plan.json, work/content-cards/cards-plan.json, work/captions/captions-plan.json, and work/shorts/shorts-plan.json.
• When several pixel operations are active on one sequence, the canonical order is cut -> color-grade -> b-roll -> captions -> content-cards -> motion-graphics. This relative order applies to every selected pair among captions, content cards, and motion graphics. Captions establish a reserved subtitle region that neither downstream operation may occupy.
• For content cards and motion graphics, the visible face and head silhouette of every primary or foreground person, speaker, presenter, interviewee, or semantically important person is a hard exclusion zone throughout the complete cue. An incidental background-only person who is not a narrative or visual focus is exempt; when classification is uncertain, protect the person. If an overlay intersects a protected face or head, reposition it first, then scale or redesign it; skip the cue if no compliant placement exists.
• B-roll adds a b-roll track and changes video pixels only; it leaves timeline, geometry, and audio untouched. Shots carry real probe/byte/SHA-256/provenance records, are normalized to timeline dimensions and exact rational FPS with the selected LUT pre-applied when color grade is active, and pass a two-stage gate: exact-candidate review, then a completed visual review bound to the final delivery. Missing or weak media becomes skipped — an approved all-skipped plan is a valid no-op. Never substitute fallback media.
• Caption cues use program time mapped from the canonical source transcript through timeline.json; they preserve source evidence and contribute a transparent PNG sequence overlay at the exact rational timeline FPS. Browser runtime assets must be local and hashed.
• Motion-graphics cues require verified video understanding, exact permissive source licenses, frozen local source/port/review hashes, deterministic native HyperFrames time, and transparent PNG sequence overlays at the exact rational timeline FPS. The shared compiler revalidates the plan and every delivery binding before rendering.
• Shorts consume the verified main delivery plus a program-time transcript. The shorts operation is present in the DAG but absent from sequences.main.operations, so derivative media under final/shorts/ never changes or re-renders the main delivery.
• Human and explicitly delegated Agent decisions are both valid when the domain receipt names the decision mode, binds reviewed artifacts by hash, and stores a non-empty rationale. Never write a fake human response for an Agent decision.
• User-facing files live in input/, review/, and final/; work/cache/ is disposable.
• Compile approved active operations with build_render_plan.py, then render delivery once with render_project.py. Timeline changes require audio filtering/encoding; -c:a copy is valid only when no active operation cuts, concatenates, retimes, or filters audio.
• video-edit-compare runs after final delivery and supports only original-vs-final-source-time.
• Default review uses stills, contact sheets, boundary reels, and short previews. Render a full-length intermediate only when a whole-program pacing decision requires it.
• Preserve compatibility adapters for current edit, looks, and cue formats until every documented consumer has migrated.
• Image-sequence overlays declare asset_type, a basename-only printf pattern, start_number, and rational fps; the compiler requires the directory, first frame, and FPS match before delivery rendering.

Architecture that spans skills

These conventions are shared and load-bearing — match them in any new skill:

• projectlib.py is the shared runtime. skills/video-understand/scripts/projectlib.py is imported directly (no package install) by every multi-skill script. It owns: project JSON validation, revision/staleness checks, map_transcript_to_timeline, build_render_plan, and durable write_json. Any new cross-skill script imports it by putting it on sys.path or running from a directory where it's co-located.
• The transcript is the shared interchange format. skills/video-understand/scripts/transcribe.py is the canonical transcriber (faster-whisper, CPU/int8, VAD, word-level). It emits transcript.json = segments[] → words[] with per-word start/end. Note it is English-only (base.en, language="en"); for other languages the caller swaps the model/lang — downstream scripts only consume the resulting JSON.
• Code is the edit. No timeline scrubbing. The edit is always text you can read, diff, and re-render: a JSON cut plan (edit_coarse.json/edit_final.json), a looks.json, an cards-plan.json, or caption cues. Beats land on the spoken word by grepping the transcript for the phrase and reading its start time.
• Human decides content; scripts do precision. Editorial calls (what to keep, which look, what a card says) are hand-authored; scripts handle boundary alignment, dead-air reclaim, varispeed, LUT baking, compositing.
• work/ for intermediates, deliverable to project root / out/. Outputs go to the passed --out dirs (durable), never a system temp dir.
• API keys come from skill-local .env, never from chat or argv. video-add-b-roll reads PEXELS_API_KEY from the environment first, then skills/video-add-b-roll/.env (gitignored). If both are blank, stop and ask the user to add it to that file. Never put a key in a command argument, plan, URL, log, review artifact, or response.
• Self-check is non-negotiable. Every skill ends by verifying its own output (re-transcribe the cut, screenshot stills, eyeball a card/skin strip) before declaring done. Don't skip it.
• Two render families:
  • ffmpeg/Python standalone operations copy audio when they do not change time. The shared delivery renderer encodes audio after cuts, concatenation, varispeed, or an audio filter; otherwise it uses -c:a copy.
  • HyperFrames (video-add-captions, video-add-content-cards, video-add-motion-graphics) renders transparent overlays at the source dimensions, then ffmpeg-composites them onto the source.
• Review before delivery. Color grade records a human or delegated agent look choice; content cards uses interview and candidate-review gates before rendering.

Windows / ffmpeg gotchas (this repo is developed on Windows)

• Drive-colon paths break ffmpeg filtergraph options that take a path (lut3d, metadata=print:file=). The pattern used throughout: run ffmpeg with cwd set to the file's folder and reference it by basename. Keep this for any new path-taking filter.
• Never use ffmpeg drawtext for labels. The fontfile drive-colon path is unreliable on Windows. Render text to a PIL PNG and overlay it (see video-color-grade). Fonts are configured at the top of the relevant script.
• Bash grep/sed one-liners in SKILL.md won't run in PowerShell. Capture ffmpeg output to a file and parse with Python instead.

Common commands

There is no root build, package manifest, or aggregate runner. Pipeline commands live inside each SKILL.md; the checks below are the repo-wide ones. Run everything from the repo root.

Sanity-check tool availability before running a pipeline: ffmpeg -version, ffprobe -version, yt-dlp --version, python -c "import faster_whisper", python -c "import PIL", node --version (>= 22, for the HyperFrames .mjs scripts in video-add-captions).

# transcribe (the shared step) — produces transcript.json + .srt
ffmpeg -y -i work/source.mp4 -ac 1 -ar 16000 work/audio16k.wav
python skills/video-understand/scripts/transcribe.py work/audio16k.wav work/transcript

Tests and checks

Tests are plain unittest modules under skills/<name>/tests/; each puts its own scripts/ and video-understand/scripts/ on sys.path, so no install or PYTHONPATH is needed. There is no discover root that covers them all — run them per skill:

python -m unittest skills/video-add-b-roll/tests/test_broll.py          # 155 tests, ~50s
python -m unittest skills/video-add-motion-graphics/tests/test_motion_graphics_plan.py
python -m unittest skills/video-cut/tests/test_inspect_bounds.py
python -m unittest skills/video-edit-compare/tests/test_make_compare.py

# one test case / one test
python -m unittest skills.video-cut.tests.test_inspect_bounds.InspectBoundsTests -v
python -m unittest skills/video-add-b-roll/tests/test_broll.py -k pexels -v

check_*.py scripts are self-contained regression harnesses with fixtures baked in — they take no project argument and print a passed line per check:

python skills/video-understand/scripts/check_protocol_extensions.py   # shared protocol
python skills/video-cut/scripts/check_project_protocol.py
python skills/video-add-captions/scripts/check_project_protocol.py
python skills/video-to-shorts/scripts/check_project_protocol.py      # wraps check_review_ui.py
node skills/video-add-captions/scripts/check_caption_style_config.mjs
node skills/video-add-captions/scripts/check_caption_interaction.mjs

Some b-roll tests shell out to ffmpeg and print UnicodeDecodeError tracebacks from subprocess reader threads on Windows (non-UTF-8 ffmpeg output). Those are noise — trust the final OK/FAILED line. As of this writing video-to-shorts/scripts/check_review_ui.py has 2 pre-existing failures (fake_render() missing video_preset); everything else above passes.

Prefer python -m unittest over pytest — the suites use unittest fixtures and mock.patch throughout, and no pytest config exists.

Repo conventions

• CLAUDE.md and AGENTS.md must stay byte-identical. They were once hardlinked but are now separate files; apply every edit to both.
• Editor UI verification. Work under ui/ must use Playwright with installed Chrome for screenshot-based visual inspection and interaction verification; typecheck/build alone are insufficient.
• work/, docs/, /tests/, and .env are gitignored. Product source includes skills/ plus the Plugin, editor runtime/UI, packaging, and launcher directories. A video project lives outside the repo and is addressed by an explicit project root, so scripts take paths, never assume cwd is the project.
• Never commit docs/ — including docs/superpowers/ — under any circumstance. It is local scratch and vendored material. Never stage it, never force-add it (git add -f), and never remove it from .gitignore. If a commit or PR would include anything under docs/, drop those paths and say so instead of committing them.
• .github/workflows/clawhub-publish.yml publishes skills/ to ClawHub (dry-run on PRs, real publish on main). A skill directory's SKILL.md frontmatter name is its published slug and slash-command trigger — renaming a directory renames the command.