claude-start

Opinionated Claude Code starter. Clone, run the install script, answer five questions, start building.

What you get

Category	What's included
Memory	CLAUDE.md with canary, error protocol, context rules — committed, teammates see the same thing
Token tools	RTK (89% avg CLI output compression) · Caveman (70% response compression) · ctx7 (live library docs)
Skills	addyosmani/agent-skills (24 lifecycle skills) · mattpocock/skills (grill-me, handoff, tdd, git-guardrails, write-a-skill) · Ponytail (anti-overengineering)
Hooks	session-start · notify · log-bash · pre-deploy-guard (lint + test gate) · privacy-guard (blocks reads of .env/ssh keys/secrets) · git-guard (reasoned block on force-push/reset --hard/git clean/rm -rf) · auto-format · keep-going · statusline (model/dir/branch/context%/rate-limit%)
Rules	.claude/rules/ — glob-scoped domain rules, plus agent-output-conventions.md (always-on: no silent truncation, explicit empty states, next-step suggestions on all agent-facing output)
Commands	/commit /pr /ship /plan /checkpoint /batch /review /verify
Workflows	.claude/workflows/ — dynamic multi-step JS scripts Claude writes, become /<name> commands (empty until you need one)
Agents	ReadOnly · BuildValidator · LogAnalyzer · Researcher · CodeReviewer · DocWriter
Logs	Decision log · error log · bash audit trail · checkpoint system

How it works day-to-day

Opening a session

The session-start hook fires automatically. It sets your terminal tab title to the current git branch and surfaces any in-progress checkpoint so you start oriented without re-reading context.

During a task

• Before anything non-trivial: /plan — scaffolds Goal / Constraints / Acceptance Criteria / Unresolved Questions in .claude/plans/
• Unfamiliar library: Claude uses npx ctx7 library <name> <query> before writing any API calls — never guesses from training data
• Files are formatted on every write — the auto-format hook runs PROJECT_FMT in the background, invisible
• All bash commands are logged to .claude/bash.log — full audit trail, no effort required
• Context getting long: /caveman halves output verbosity. Before stopping: /checkpoint saves exact state

Finishing a task

• Every completed task ends with [Canary:PROJECT_NAME:TASK_NAME] — the keep-going hook checks for it and nudges Claude to continue if it's missing
• /handoff instead of /compact — precise surgical session doc, not lossy compression

Shipping

/ship

That's it. Runs lint → build → test → commits → opens a PR. Stops at the first failure.

Or deploy directly — pre-deploy-guard intercepts the deploy command and runs your lint and test suite first. Blocks with a reason if anything fails.

Code review

/review

Invokes the CodeReviewer agent on your current diff. Returns a structured report with severity-tagged issues. You decide what to act on.

Working with teammates

Commit .claude/ — your teammates get the same hooks, agents, and commands automatically. They run through ONBOARDING.md to set up their local settings and understand what everything does.

Requirements

• Node.js 18+
• Rust + Cargo (for RTK)
• Claude Code installed and authenticated
• git, gh (GitHub CLI) for /commit, /pr, /ship

Install

macOS / Linux / WSL:

git clone https://github.com/moneytosms/claude-start my-project
cd my-project
chmod +x install.sh && ./install.sh

Windows:

git clone https://github.com/moneytosms/claude-start my-project
cd my-project
.\install.ps1

The script installs all tooling, wipes the template git history, inits a fresh repo, then opens Claude to ask you five questions: project name, description, runtime, formatter, and canary codename. That's the only interactive step.

Adopting into an existing project

Already have a repo with its own .git and CLAUDE.md? Don't use install.sh — it wipes git history. Use install-existing instead: it never touches .git, backs up your existing .claude/ before replacing it, and asks Claude to reconcile anything from the backup plus merge CLAUDE.md knowledge into yours (your file stays the base).

macOS / Linux / WSL:

git clone https://github.com/moneytosms/claude-start /tmp/claude-start
cd /tmp/claude-start
chmod +x install-existing.sh
./install-existing.sh /path/to/your/existing/project

Windows (PowerShell):

git clone https://github.com/moneytosms/claude-start C:\temp\claude-start
cd C:\temp\claude-start
.\install-existing.ps1 -Target C:\path\to\your\existing\project

What it does:

1. If you have an existing .claude/, backs it up to .claude.backup-<timestamp>/, then writes the template's .claude/ (hooks, agents, commands, rules) in its place
2. Same for .mcp.json — backed up to .mcp.json.backup if present, then overwritten
3. Creates .claude/settings.local.json from the example if missing
4. chmod +x on hooks (Linux/WSL/macOS directly; on Windows, run that step yourself in WSL/Git Bash afterward)
5. If you have no CLAUDE.md, copies the template's in as a starting point
6. Opens Claude in your project with one prompt that does three things: merges useful template CLAUDE.md sections into yours (your file is the base, nothing overwritten or restructured); detects your stack (package.json, pyproject.toml, Cargo.toml, go.mod, Gemfile, pom.xml/build.gradle, *.csproj, etc.) and fills the RTK wrappers line under ## Stack with the real rtk subcommands for what's actually present; and reconciles the .claude.backup-*/ — ports back anything project-specific (custom hooks/agents/commands/rules) you had before, then deletes the backup and the reference copy

Nothing is committed for you — review the diff (git status, git diff) and commit when happy.

After install

1. Fill in the rest of CLAUDE.md as the project takes shape — Test, Lint, Build, Deploy commands
2. Create .claude/rules/ files as major directories emerge — see rules/example.md for the format
3. Teammates: run through ONBOARDING.md to get their local settings configured
4. Run /ponytail-audit once the codebase has real code — finds complexity to cut
5. Check rtk gain after a few sessions — shows your actual token savings

What's not included

• MCP servers — add to .mcp.json (team-shared, committed) or settings.local.json (personal)
• CI/CD — /ship handles local → remote; pipeline config is project-specific
• Issue tracker — add mattpocock's triage skills if needed

License

MIT — see LICENSE.