Claude Code Starter Pack — Documentation skill for Claude Code

Claude Code Starter Pack

A [Claude Code skill](https://docs.anthropic.com/en/docs/claude-code) that scaffolds a documentation system into a project — new or existing — to help a long-lived AI agent stay grounded across sessions: not re-deciding settled questions, not over-claiming what works, and not editing its own north star.

It comes out of a few long-running AI-agent projects. The reasoning behind each file is in [`starter-pack/references/anti-drift-playbook.md`](starter-pack/references/anti-drift-playbook.md).

The idea

A north-star document captures intent; drift happens in the decisions and invariants made while pursuing it. So ground those separately, re-inject them each session, and make the immutable ones read-only to the agent.

Install

git clone https://github.com/Ayush-sk-Pathak/claude-code-starter-pack.git
cp -R claude-code-starter-pack/starter-pack ~/.claude/skills/starter-pack

Then, in any project, tell Claude Code:

/starter-pack

(or just say "set up a new agent project"). It interviews you — project one-liner, **your vision in your own words** (hard gate: nothing lands in `vision.md` without your explicit line-by-line approval), stack, branch model — then scaffolds everything, commits, and installs the enforcement layer. On an existing repo it runs a gap analysis and adds only what's missing; it never overwrites an existing doc.

What gets scaffolded

Documents are split by **what drifts** — each layer has one owner, one change discipline, and one kind of drift it prevents:

File	Layer	Discipline
vision.md	Intent — the north star, in the user's own words	User-only; chmod 444 + pre-commit rejected
CLAUDE.md	Laws — Karpathy's four rules verbatim + standing protocols	Rare, deliberate; protected
docs/architecture.md	Design — blueprint of record	Deliberate; protected
docs/STRATEGY-LEDGER.md	Decisions & invariants — settled calls, rejected