CutForge: agent context

This file is auto-loaded by OpenAI Codex, Google Antigravity, GitHub Copilot CLI, Cursor (when in the agent context chain) and other Agent Skills runtimes. It is the equivalent of CLAUDE.md for non-Claude surfaces.

What this plugin is

CutForge is an open-source video creation and editing plugin: from raw footage to final cut. It turns a client's photos, clips, voice notes, documents and brand assets into a finished, verified, story-led video: family and milestone films, tributes, weddings, brand films, explainers, ads, and pharma or BFSI videos with India-first compliance. 13 skills · 4 reviewer agents · a tested engine of 53 tools (Remotion, ffmpeg, Python) · 0 hooks · 0 MCP servers.

Supported surfaces (v1.1.2): Claude Code (CLI + IDE extensions), Anthropic Cowork, OpenAI Codex (CLI + IDE + App), Cursor 2.5+, GitHub Copilot CLI, Google Antigravity 2.0 (CLI + IDE), Hermes Agent, OpenClaw, Grok (xAI Build CLI, native .grok-plugin/ pair).

Core principles

1. Real footage first; AI only where no camera was, and never on real faces.
2. Every time is derived, never typed. film/timeline.py turns measured narration into every shot's length, so picture, captions and sound always agree.
3. The project folder is the memory. Facts, decisions, sources, spend and notes live in files, not in chat.
4. Never guess identity. People, objects, rituals, dates and quotes are verified, or asked in one batched round.
5. Verify before saying done. Stills, audio QC, frame counts and decode checks run before anyone watches.

How to use it as an agent

1. Start with the cutforge skill. It routes the phases (brief, archive, story, look test, storyboard, voice, sound, art, build, QC, delivery, revisions) to the specialist skills.
2. Scaffold a project with python <plugin>/engine/new_project.py <folder> --title "<title>" [--demo], then npm install and python tools/setup_local.py inside it. <plugin> is this repository's folder (Claude hosts set ${CLAUDE_PLUGIN_ROOT} to it).
3. Project tools run inside the project: python film/timeline.py, python tools/qc/stills.py <chapter>, bash tools/render/pipeline.sh <chapter>, python tools/mix/film_mix.py, bash tools/render/assemble.sh. The project's own CLAUDE.md lists the full pipeline.
4. Paid calls are quoted first. A voice or AI-art tool run without --go prints a quote and exits 4, spending nothing: the price comes from tools/price_book.py, recorded from the provider's own page, never from memory. Run it again with --go --quoted-providers <the quoted providers> only after the client's typed go. Tools stop at config.budget_inr; every paid call is logged to spend.md. Keys come from environment variables or a key file, never from the chat.

Canonical entry points

Files in this repo

• skills/<name>/SKILL.md: 13 Agent Skills, each with references/ for detail.
• agents/<name>.md: 4 reviewer agents (story-editor, film-qc-reviewer, compliance-reviewer, archive-scout).
• engine/new_project.py + engine/template/: the project scaffold and every tool (render, mix, grade, QC, archive, VFX, voice, delivery). engine/templates/: the project memory files.
• docs/lessons-learned.md: every mistake from a real 20-minute production and the rule that prevents it.
• docs/case-study.md: that production, anonymised. docs/compliance-and-licensing.md: licences, platform specs and Indian ad compliance with sources. docs/USER-GUIDE.md: a walkthrough.

Cross-platform notes

• Skills are portable across all supported surfaces (Agent Skills open standard).
• Subagents are Claude-only as agents/*.md; on other surfaces run the same checks by asking for an independent review with the agent file's instructions.
• The engine needs Node 22 or newer, Python 3.10 or newer (3.10 reached end of life on 2026-10-01 and still runs; use 3.11 or newer where you can) and ffmpeg on PATH. Windows: run .sh files in Git Bash and set PYTHONUTF8=1.

Identity / authority

Built and maintained by Indranil Banerjee (https://indranil.in). MIT-licensed. No telemetry. Part of the Neelverse plugin suite.