Home

project-init

Scaffolder that drops agentic-development infrastructure into any project: agent instructions, memory, deterministic guard hooks, GitHub workflow enforcement, and a Claude Code plugin — ready from day one for Claude Code, with opt-in support for Codex, Gemini CLI, and Ollama-based agents.

What it does

One command → structured layout inside your project:

<your-project>/
├── AGENTS.md                # canonical agent instructions (Linux Foundation standard)
├── CLAUDE.md / GEMINI.md    # thin redirects to AGENTS.md
├── CONTRIBUTING.md / SECURITY.md / .github/CODEOWNERS
├── .env.example             # documented env pattern (.env gitignored)
└── .claude/
    ├── config.yaml          # options chosen at init + scaffold record (drives `upgrade`)
    ├── settings.json        # plugin marketplace wiring (+ hooks with --no-plugin)
    ├── hooks/dag_workflow.py, scripts/, rules/, docs/
    ├── memory/MEMORY.md     # grep-able memory index
    └── vault/               # Obsidian vault (humans)

Scaffolds are plugin-first: the project-init-workflow plugin (hosted in this repo's marketplace) provides the shared skills and guard hooks with auto-update. --no-plugin copies everything locally instead (offline / no-trust environments).

Re-running is safe — it reconciles and never overwrites memory or vault notes. project-init upgrade [--apply] re-renders from the recorded config with a drift report; local edits become .new siblings, never overwritten.

Presets

Preset	What you get
obsidian-only	Base scaffold + Obsidian vault for humans
obsidian-graphify	Above + Graphify code knowledge graph for agents (ADR-009)

(The former obsidian-lightrag preset was removed — upgrade auto-migrates recorded projects to Graphify.)

Install

From PyPI (preferred once the first release is published — ADR-011):

uv tool install project-init     # or one-off: uvx project-init .

Full setup from git (adds the /project-init slash command for Claude Code):

curl -sSL https://raw.githubusercontent.com/VytCepas/project-init/main/install.sh | bash

Key flags

Flag	Effect
--agents claude,codex,gemini,ollama	native wiring overlays per agent (Claude default)
--license mit|apache-2.0|proprietary + --owner	LICENSE, CODEOWNERS, SECURITY contact
--devcontainer / --mise / --vscode	opt-in environment overlays
--no-plugin	copy hooks/skills locally instead of the plugin
--strict	fail on any unrendered template placeholder

Safety model (ADR-007 / ADR-012)

Agent hooks (commit gate, GitHub command guard, prod-safety guard) are fast-feedback guardrails. The enforcement boundary is git hooks + CI (validate-pr, secret scan) — run .claude/scripts/install_hooks.sh once per clone — plus credential separation: agent sessions never hold production credentials.

Docs

• README — full user docs
• ADRs — every architectural decision
• Using project-init guide