open-edit AGENTS.md

AGENTS.md

This repo is a video creation and editing system powered by veed-engine-cli (the veed render engine). It is not limited to captions: agents can edit, cut, and reframe footage; layer motion graphics and visual elements (shapes, images, animations — not just text); turn slides or websites into video; capture web pages; and compose rich stylized output over real footage.

Video is not always required. Work from stills, slides, websites, generated imagery, audio-only sources, or pure motion graphics when that fits the ask — source footage is optional, not assumed.

Use the full toolbox. Reach for any video or image generating service, and any available MCP server, when it helps — don't limit yourself to the local engine and recipe path.

Be creative and fluid. Try new ideas and avenues. Prototype, explore alternatives, and follow promising directions rather than collapsing every ask onto the captioned-recipe default. Match the ask; invent when the brief is open.

One common path is stylized captioned video via the open-edit skill (.claude/skills/open-edit/SKILL.md) The default (recipe-backed) caption run is scripted end to end — recipes are compiled code (refs/html/<id>/recipe.ts), zero tokens at run time; subagents remain only for the opt-in vision analysis (run when the user asks to refine the style); the CREATIVE path face-1 (the user brought their own reference/brand/concept) is authored INLINE by the orchestrator, and creative iteration on a delivered result is REMIXED inline. Treat captions as one capability among many — match the ask, and use the full engine surface when the work calls for edits, motion graphics, or other compositions.

To run a video

When the user says "run the open-edit skill on <video.mp4>", follow .claude/skills/open-edit/SKILL.md exactly (the FAST PATH): preflight (automatically prepare the workspace-local runtime; ask before global installs or updates) → prep (transcript from the recorded provider — npx @veedstudio/openedit-cli transcribe, with --provider veed for the hosted route — then npx @veedstudio/openedit-cli prep: meta.json + word-timings.json from the transcript's real per-word times + base frames) → sample ONE style (npx @veedstudio/openedit-cli sample-style, facet-scored, seeded, zero tokens → style.json; the pool is the recipes-only runtime index refs/tags.json, so the draw is always recipe-backed) → design + render (npx @veedstudio/openedit-cli generate-recipe --run runs/<key> --record, a SCRIPT driving the full gate chain: the compiled recipe emits the .wv document → lint → --verify with the mechanical ladder fix loop → records out.silent.mp4 → probe-qa frame QA) → mux audio (npx @veedstudio/openedit-cli mux-audio). Deliverable → runs/<key>/final/out.mp4. The CREATIVE PASS routes on the SHAPE of the ask (the user never learns recipes exist): face-1 — the prompt arrives WITH the user's own reference/brand/concept → the orchestrator authors design+render INLINE per the brief (no subagent: nothing to orphan in headless runs, and the design stays in the session's context so follow-up tweaks iterate instantly); the USER'S materials are the design authority (they get looked at); 1-2 nearest recipe sheets ride along as craft substrate (mechanics only) + a DIRECTION (content/mood/placement, never fonts/palette/device). face-2 — iteration on a delivered result: ANY creative input ("make it more interesting", a mood, an aesthetic change) → REMIX, inline per the brief's REMIX MODE (donor sheets from the index, fresh runs/<key>-remix, same gates); "show me more options / N versions" (zero creative input) → RE-ROLL (N seeded draws through the script path); defect repairs → fix at the source + re-run the gates (agent-authored templates are patched directly; recipe outputs are script-owned — regenerate via inputs, the refine path, or a --module copy, never hand-edits). The analyse step (nameless bg vision subagent → analysis.json) is OPT-IN: run it only when the user asks to refine the style, then re-run design+render as the from-scratch inline pass (compiled recipes ignore analysis.json; the pass composes from it) and re-mux — analysis.json existing IS the fast-path/refinement switch.