All skills
latent-spaces avatar

/nanoclip-cutting-room

@c78d0da

Turn long footage (podcast, interview, stream, recording) into ready-to-post vertical shorts. Use when the user points at a long video and wants short clips β€” "cut this into shorts", "make clips from this recording", "turn this podcast into verticals". Needs the NanoClip CLI (analysis) and HyperFrames (rendering). NOT for editing an existing short, generating footage, or plain transcription.

Use this Skill: https://skilld.dev/gh/latent-spaces/nanoclip-cutting-room/nanoclip-cutting-room

This session only. Nothing lands on disk.

SKILL.md

β‰ˆ105 tokens always: the name and description. β‰ˆ917 when used: this file.

nanoclip-cutting-room β€” router

What this is

NanoClip is the intelligence (transcript, diarization, face clustering, active-speaker detection, scene cuts). HyperFrames is the rendering engine (HTML compositions β†’ 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 β€” 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 β€” 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 β€” node/ffmpeg, installs the NanoClip CLI
             from npm when missing, warms the pinned HyperFrames version
1 CARDS      pick video Β· language Β· key check (conditional)
2 INTAKE     mezzanine rule β†’ upload β†’ server quote
3 SPEND GATE card with exact USD β†’ start --approve (CLI refuses overruns)
4 THE SCREEN opens in browser; watcher waits β€” zero agent tokens
5 DATA LANDS digest.mjs unifies payloads; cast narrated in chat
6 CUT        editorial pass (references/editorial.md) β†’ plan.json β†’ scaffold
7 PREVIEW    the clips embed in the Screen (hyperframes play behind them)
8 ITERATE    all steering happens in chat
9 SHIP       Finisher gate β†’ local render β†’ vertical .mp4s Β· 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 β€” 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 β€” 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 β€” 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 β€” no in-browser editing tools.

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

Source: SKILL.md on GitHub

1 warning11d3 checks Β· Risk SAFE
  • Gen Agent Trust Hub11d

    The skill is a comprehensive video editing suite for creating vertical shorts. It uses the NanoClip service for analysis and HyperFrames for rendering. Security analysis indicates excellent defensive practices, including a 'token firewall' to prevent context exhaustion, strict HTML escaping of untrusted transcript data, and a policy of never handling user API keys. External dependencies are restricted to vendor-owned or well-known packages from official registries.

  • Socket11d

    3 alerts: gptAnomaly

  • Snyk11d

    Risk: LOW Β· No issues

Signed by skilld at c78d0da. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub last month.

Steadyupdated last month

README badge

README badge for latent-spaces/nanoclip-cutting-room