OpenClip — the agent-orchestrated video editing harness

# skills → Claude Code, Codex (tested), Cursor + any skills-protocol agent npx skills add Q00/openclip # the oc CLI (ffmpeg + Python 3.11+ required; pin to a release tag) uv tool install "git+https://github.com/Q00/openclip@v0.1.0"

#The shape of a run

flow manifest ──▶ orchestrator agent │ read open steering (oc steer) ┌─────────────┼─────────────┐ parallel wave ▼ ▼ ▼ worker worker worker one subagent per unit └─────────────┼─────────────┘ (chunk · section · clip) ▼ merge JSON surface to human ◀── steer ──┐ ▼ approved │ render / cut │ ▼ │ oc-verifier ── needs-fix ───┘ re-dispatch the unit ▼ confirmed checkpoint ──▶ next stage

One orchestrator, parallel workers, a human checkpoint, and an independent verifier — only a confirmed verdict advances.

#Four flows

`flow1` Cut-edit

Proxy → parallel STT → a cut-editing debate: proposers argue through filler / pacing / narrative lenses, a judge reconciles → cut-edited original + subtitles.

`flow2` Shorts

One long video → hook mining across transcript sections → captioned 9:16 shorts, ends snapped to the payoff's last word — plus a thumbnail per short.

`flow3` Assemble

Weave N source videos into one longform, then mine its hook moments into shorts.

`flow4` Thumbnails

Hook-matched thumbnails: the most representative frame with a burned headline, and/or a gpt-image design.

#Why it's different

Debate beats a single voice. Cuts aren't decided by one pass — competing editorial lenses argue, a judge reconciles, and the human can overrule any of it.

A "done" is a claim, not proof. Every render passes a mechanical evidence gate (duration, aspect, silent-audio, black-tail, SRT validity) and an independent adversarial verifier. Only confirmed advances.

The human is the director. oc steer --note "cut the intro harder" mid-run; workers honor it over their own judgment. Nothing creative auto-converges past you.

Interruption-proof. Renders are keyed in an event ledger — a re-run skips finished work (resumed: true) instead of re-billing Whisper and gpt-image.

It grows. Agents author missing tools into a git-tracked toolbox; a separate auditor gate promotes them to shared only after a scrubbed-env re-verify and deny-list scan.

#Integrating as an agent

If you are an AI agent evaluating or driving this harness: start at llms.txt (repo map + invariants), then AGENT_GUIDE.md (flow routing, fan-out widths, verification rules). Every oc subcommand prints exactly one JSON object; errors print {"error": "...", "type": "..."} and exit non-zero.

#The shape of a run

#Four flows

flow1 Cut-edit

flow2 Shorts

flow3 Assemble