context-forge

Auto harness engineering for AI coding agents. Discuss your project, curate battle-tested know-how, and bootstrap a fully context-engineered repository — for Claude Code, Cursor, Copilot, and beyond.

🇰🇷 한국어 · Catalog · Discussions · Contributing

---

What is context-forge?

context-forge is a meta repository that turns scattered AI-agent know-how — CLAUDE.md patterns, AGENTS.md standards, .cursorrules, agent skills, MCP servers, spec-driven workflows, prompt libraries, and convention guides — into a single curated catalog that you can apply to a brand-new project in one go.

You describe what you're building (web app, Unity game, MCP server, CLI tool…) through a guided discussion. context-forge picks the matching know-how from the catalog, creates a new GitHub repository, and writes everything in: CLAUDE.md, AGENTS.md, README.md, docs/spec.md, .claude/skills/, .claude/agents/, .claude/rules/, boilerplate, and more.

Think of it as create-react-app for AI-augmented codebases — except instead of a fixed template, it's a living catalog crowd-curated from the best agentic-engineering repos on GitHub.

Why?

The current state of AI coding agents (Claude Code, Cursor, Copilot, Codex, Aider, Windsurf, …) demands serious context engineering before useful work begins. Every project re-solves the same boilerplate:

• Which CLAUDE.md patterns apply here?
• Which skills / subagents / hooks should be installed?
• Which spec-driven workflow fits this team?
• Which MCP servers do we need?
• Which language conventions, code-review checklists, commit standards?

context-forge collapses that hours-long ritual into a guided 5-minute conversation. Then you cd into the new repo and start coding.

How it works

[Forked know-how repos]                glowElephant/<fork>...
        ↓
[Catalog]   ───────►   /context-forge slash command
        ↓                  │
[Discussion]               │ Brief fixed questions
        ↓                  │ Free-form goal/constraint chat
[Match + Confirm]   ◄──────┘
        ↓
[gh repo create user/<project>]   →   New harness-engineered repo
                                        ├── CLAUDE.md / AGENTS.md
                                        ├── README.md
                                        ├── docs/spec.md
                                        ├── docs/know-how/
                                        ├── .claude/skills,agents,rules/
                                        └── boilerplate

Catalog scope (v1)

The v1 catalog is intentionally narrow — focused on the AI-agent harness itself:

Category	Examples
Claude Code core	anthropics/skills, obra/superpowers, garrytan/gstack
Cursor / Copilot rules	PatrickJS/awesome-cursorrules, github/awesome-copilot
Spec-driven development	github/spec-kit, Fission-AI/OpenSpec
Multi-agent orchestration	wshobson/agents, ruvnet/ruflo
MCP infrastructure	modelcontextprotocol/servers, punkpeye/awesome-mcp-servers
Standards	agentsmd/agents.md, google-labs-code/design.md
Prompt engineering	dair-ai/Prompt-Engineering-Guide, f/prompts.chat

v2 will broaden to domain conventions (game engines, mobile, web, backend, ML, security, documentation). See docs/specs/2026-04-28-fork-candidates.md for the full curated list.

Status

✅ v1.1.0 released (release notes)

• Design specification
• Candidate repositories curated (67 v1 + 210 v2)
• Forking the v1 catalog sources (82 active after first monthly adoption)
• /context-forge slash command implementation
• Catalog frontmatter for every source (82/82) — manual v1 seed (15) + auto-seeded via scripts/seed_frontmatter.py (67)
• Weekly fork sync + link-check (docs/fork-sync-setup.md)
• Monthly borderline rescore + quarterly full rescore (5-factor rubric)
• Monthly auto-discovery job → Issue triage (automation.py discover + monthly-rediscovery.yml)
• Topic tag context-forge-source on all forks (browse)
• End-to-end smoke test — Phase 0~6 dogfood-validated 2026-05-27 (2 bugs found & fixed: YAML quote leak in auto-seeded category, git branch -M main missing for Windows). Fresh-session test in real Claude Code still recommended for full Skill-tool invocation flow.
• Automated catalog extraction from forked sources (v1.5)
• v2 domain expansion (game/mobile/web/backend/ML/security)

Track progress in docs/specs/ and docs/plans/.

Quick start

Prerequisites

#	Requirement	Check
1	gh CLI authenticated	gh auth status
2	context-forge cloned locally	git clone https://github.com/glowElephant/context-forge ~/code/context-forge
3	CONTEXT_FORGE_PATH env var pointing at the clone (optional — skill asks once per session if missing)	see below

The slash command itself runs validate-catalog.sh at startup and refuses to proceed if anything is wrong, so the catalog state is verified for you.

Set CONTEXT_FORGE_PATH

macOS / Linux:

echo 'export CONTEXT_FORGE_PATH=~/code/context-forge' >> ~/.bashrc   # or ~/.zshrc
source ~/.bashrc

Windows (PowerShell, persist across all shells):

[Environment]::SetEnvironmentVariable('CONTEXT_FORGE_PATH', 'C:\Git\context-forge', 'User')
# 새 터미널/세션부터 적용

Claude Code only (works across Mac/Linux/Windows, syncs with .claude/settings.json):

{
  "env": {
    "CONTEXT_FORGE_PATH": "C:\\Git\\context-forge"
  }
}

Run

In any directory:

/context-forge

Answer a few questions about your project. context-forge will:

1. Match catalog know-how (82+ sources) to your stack and goals
2. Create a new GitHub repository
3. Populate it with curated CLAUDE.md, skills, rules, docs
4. Hand off — cd <new-repo> && claude to start coding

v1.1.0 — weekly fork sync + monthly auto-discovery active. End-to-end smoke test still pending; report issues via Discussions.

Contributing

context-forge is a community catalog. Add your favorite repository or your own:

• 💬 GitHub Discussions — propose a source repo to fork (open a discussion)
• 🛠 Pull request — directly add an entry to sources/index.json
• 📋 Scoring rubric — see docs/scoring.md

All sources are evaluated on popularity, activity, real-world reviews, content quality, and maintainer reliability before promotion to Tier 1 (auto-fork).

Why "harness engineering"?

The agentic coding community has converged on the term harness for the layer that wraps an AI model — system prompt, tools, memory, hooks, context. Harness engineering is the discipline of designing that layer well. context-forge is the meta tool that does this for you, automatically, before you write a single line of code.

Related work

context-forge stands on the shoulders of these projects:

• obra/superpowers — agentic skills framework
• garrytan/gstack — Claude Code 23-tool stack
• github/spec-kit — spec-driven development
• agentsmd/agents.md — AGENTS.md open standard
• coleam00/context-engineering-intro — context engineering primer

License

MIT © glowElephant

---

Keywords: claude code · claude skills · agent skills · cursor rules · copilot instructions · agents.md · claude.md · context engineering · spec-driven development · mcp servers · harness engineering · agentic coding · ai coding assistant · prompt engineering · meta repository · auto bootstrap