nanoclip-cutting-room — router

What this is

NanoClip is the intelligence (transcript, diarization, face clustering, active-speaker detection, scene cuts). HyperFrames is the rendering engine (HTML compositions → render). This skill is the editor between them. The promise: point it at a recording, approve a quote measured in cents, get three shorts you'd actually post, ask for more in plain language.

Where things live

• Skill folder — the directory holding this SKILL.md (typically ~/.claude/skills/nanoclip-cutting-room). Every node scripts/<x>.mjs in this file and in references/ is relative to it: run node <skill folder>/scripts/<x>.mjs from wherever you are. Never copy the scripts next to the footage.
• Working folder — the folder holding the user's footage (usually the cwd). All run output goes to <working folder>/cutting-room/ (plan.json, data, drafts, renders).

The flow

0 PREFLIGHT  node scripts/preflight.mjs — node/ffmpeg, installs the NanoClip CLI
             from npm when missing, warms the pinned HyperFrames version
1 CARDS      pick video · language · key check (conditional)
2 INTAKE     mezzanine rule → upload → server quote
3 SPEND GATE card with exact USD → start --approve (CLI refuses overruns)
4 THE SCREEN opens in browser; watcher waits — zero agent tokens
5 DATA LANDS digest.mjs unifies payloads; cast narrated in chat
6 CUT        editorial pass (references/editorial.md) → plan.json → scaffold
7 PREVIEW    the clips embed in the Screen (hyperframes play behind them)
8 ITERATE    all steering happens in chat
9 SHIP       Finisher gate → local render → vertical .mp4s · prefs updated

Run the preflight before anything else in a fresh session: it verifies node and ffmpeg/ffprobe, installs nanoclip from npm when the CLI is missing, reports its auth state (auth itself is the user's step — see Non-negotiables), and warms the pinned hyperframes version so the first render never stalls on a download.

Stage playbooks (load lazily, one at a time)

Stage	Reference
Cards + upload + spend gate	references/intake.md
The Screen (server, SSE, catalog)	references/screen.md
Digest (fusion join, cast naming, thumbs, gap frames)	references/digest.md
Editorial pass (what makes a clip)	references/editorial.md
Compose (extract, reframe, scaffold, captions, players)	references/compose.md
Iterate & ship (chat loop, Finisher, render)	references/iterate.md

Every stage above is built and has run end-to-end on real footage — intake through delivery renders. Do not improvise around a playbook; when something is out of scope, say so.

Non-negotiables (locked product decisions)

• Money moves only through the spend gate: exact USD per line item, dispatch with --approve <exact>, the CLI refuses live overruns. Every auto-decision gets one chat line.
• Zero agent tokens while analyses cook: the watcher uses get --wait per analysis (status returns a very large payload — rare fallback only).
• plan.json (schema cutting-room/plan@2) is the contract: pointers, never payloads; caption block singular, palette plural; Claude edits plan and compositions, the screen never does.
• The agent never handles the NanoClip key; the user runs nanoclip auth login themselves.
• Chat is the editor — no in-browser editing tools.

fixtures/README.md documents the synthetic demo payloads the tests and the replay demo run on.