Demo

A real /claude-setup run on the acme-saas Node monorepo (pnpm + turbo + Next.js + Hono + Drizzle). The skill detects the stack from lockfiles and package.json workspaces, scrapes hard rules from the README, and writes a path-scoped .claude/ tree:

Each rules/*.md is path-scoped (frontmatter paths:) so it only enters context when you open a file under that subtree. skills/release/ and skills/preview-deploy/ are wired to commands that actually exist in this repo (changeset + Vercel) — no TODO bodies, no docs-example boilerplate.

See examples/acme-saas.md for the full file-by-file breakdown including which README phrase encoded as which rule.

Install

Quickest — zero install, runs once:

npx -y @claude-setup-skill/install

Global — keeps a claude-setup-install command around to rerun anytime:

npm i -g @claude-setup-skill/install && claude-setup-install

From source — for contributing or live-editing the skill:

git clone https://github.com/sir-ad/claude-setup-skill ~/claude-setup-skill
cd ~/claude-setup-skill && ./install.sh

All three populate ~/.claude/skills/claude-setup/. The npm methods copy files; from-source symlinks them so edits in your clone are immediately live.

Use

In any project:

cd ~/some-new-project
claude
> /claude-setup

Optional flags:

| Flag | Effect | |---|---| | --minimal | only CLAUDE.md + .claude/settings.json | | --skip-skills | skip skill generation | | --dry-run | print the plan, write nothing |

The skill walks five phases:

1. Detect — reads README, lockfiles, manifests, existing .claude/.
2. Match template — picks one of templates/<stack>.md as the reference.
3. Plan — prints what it'll write; asks before clobbering.
4. Generate — writes files with project-specific substitutions.
5. Verify — prints final tree, line counts, and one verification step.

What it generates (when applicable)

| File | When | |---|---| | CLAUDE.md | always | | .worktreeinclude | always | | .claude/settings.json | always | | .claude/rules/<n>.md | per significant subdir with distinct conventions | | .claude/agents/<lang>-reviewer.md | always (read-only review subagent) | | .claude/skills/release/ | if version file + CHANGELOG.md exist | | .claude/skills/<other>/ | per detected workflow (db-migrate, preview-deploy, etc.) |

What it never generates:

• .claude/output-styles/ — no use-case-off-the-shelf
• .claude/commands/ — Anthropic docs recommend skills for new workflows
• .claude/agent-memory/ — auto-populated when subagents with memory: frontmatter run
• Empty folders
• Skills with TODO bodies
• Hard rules invented from thin air (only pulled from README / RFCs / ADRs)

Stacks supported

| Stack | Template | |---|---| | Rust workspace | templates/rust-workspace.md | | Rust single crate | templates/rust-single.md | | Node monorepo (npm / pnpm / yarn workspaces, turbo, nx, lerna) | templates/node-monorepo.md | | Node single (Next.js, Vite, Astro, Express, ...) | templates/node-single.md | | Python (Poetry, uv, pip, hatch, rye) | templates/python.md | | Go | templates/go.md | | Anything else | templates/generic.md |

To add a stack: drop a new templates/<name>.md, then add a detection branch in SKILL.md Phase 2.

Updating

Pull this repo, that's it — the symlink keeps the skill live:

cd ~/claude-setup-skill
git pull

Uninstalling

./install.sh --uninstall

Removes the symlink. Doesn't touch this repo or any project's .claude/ directories you've generated.

Why a skill, not a CLI

A skill is one markdown file Claude reads at runtime. The "context detection" is just Claude reading your project, which is already what Claude does best. A standalone CLI would either need a fixed templating engine (less context-aware) or shell out to the Claude API (more setup). Skill = zero install, maximum context-awareness, one file to maintain.

License

MIT. See LICENSE.