ai-project-init
Initialize any project for Claude Code. Three subagents author AGENTS.md, MANIFEST.md, .claude/rules/, .claude/commands/, .claude/agents/, .mcp.json, and settings.json from your project's actual evidence — no pre-built packs, no template vocabulary, no fixed taxonomy. Works on UI design systems, backend APIs, libraries, mobile apps, CLIs, ML pipelines, and anything in between.
What it does
ai-project-init is a Claude Code skill that bootstraps an agent system tailored to the project it actually lands on. The skill itself ships no opinion about what kind of project that is — there are no pre-built "UI" or "backend" or "library" packs, no atomic-design or layered-architecture vocabulary baked in. Whatever lands in the user's .claude/ is authored at install time from evidence the project-analyzer subagent extracted from the real codebase.
How it works
Three subagents and a thin orchestrator:
-
project-analyzer (read-only). Walks package.json / pyproject.toml / Cargo.toml / go.mod / etc., framework configs, README, sample source files, existing AI configs. Emits one structured analysis JSON describing stack, conventions, domain vocabulary, recommended rule topics, recommended slash commands, recommended agents, and a project-specific MANIFEST schema.
-
rules-author (writes only into .claude/.staging/). Reads the analysis JSON. Authors AGENTS.md, MANIFEST.md, .claude/rules/<topic>.md, .claude/commands/<name>.md, .claude/agents/<name>.md from scratch, citing real paths, real aliases, real packages, real domain words. No template substitution. No vocabulary the analyzer didn't surface.
-
realism-verifier (read-only). Checks every staged file: paths exist (test -e), import aliases resolve through tsconfig.json, packages exist in the project's manifest, slash-command targets are real folders, no {{VAR}} placeholders, no contamination tokens leaked in. Returns pass / fail per file with corrections JSON. The orchestrator gates promotion on a clean pass.
If verification fails, the orchestrator hands the corrections JSON back to rules-author for one or two retries. After the last failure, the install aborts; the project's existing files are untouched (everything was staged).
The orchestrator phases:
0. Mode + marker check
1. project-analyzer → /tmp/ai-init/analysis.json
2. Plan review with user
3. Copy generic shared assets to .claude/.staging/
4. rules-author → fills out .claude/.staging/
5. realism-verifier → /tmp/ai-init/verification.json (gate)
6. validate.py → mechanical structural checks (gate)
7. Atomic promote staging to project
8. (optional) seed_manifest.py for retrofit MANIFEST seeding
Install
Via Claude Code plugin marketplace (recommended)
# In Claude Code:
/plugin marketplace add ivanmeda993/ai-project-init
/plugin install ai-project-init@ai-project-init
Then in any project root: ask Claude "init this project for AI" / "namesti claude rules" / "setup AGENTS.md".
Via curl one-liner
curl -sSL https://raw.githubusercontent.com/ivanmeda993/ai-project-init/main/install.sh | bash
Manual
git clone https://github.com/ivanmeda993/ai-project-init.git
mkdir -p ~/.claude/skills
cp -r ai-project-init/skills/ai-project-init ~/.claude/skills/
Usage
In any project, any of the following triggers the skill:
- "init this project for AI"
- "namesti claude rules ovde"
- "setup AGENTS.md and MANIFEST"
- "init claude config for this backend / mobile / library / pipeline"
- "scaffold rules for this repo"
The skill spawns the analyzer, shows you a digest of what it found and what it would author, you approve or refine, then it runs through authoring + verification + promotion.
Project types
The skill is project-type-agnostic. The analyzer reads what's there and the rules-author authors rules that fit. Verified-by-design coverage:
- TypeScript / JavaScript projects (Node, Bun, Deno) — frontend, backend, library, mobile (Expo / React Native).
- Python projects (Django, FastAPI, Flask, ML pipelines, CLI tools).
- Go, Rust, Swift, Kotlin, Java, Ruby, PHP, .NET — the analyzer reads each language's manifest format.
- Monorepos (Turborepo, Nx, pnpm/Yarn workspaces, Cargo workspaces, Go workspaces).
- Hybrid projects (UI + backend, library + CLI, etc.).
If the analyzer can't read a manifest format, it surfaces that under open_questions and the user can answer.
What it does NOT do
