workflow

my-dotclaude

My Claude Code setup, version-controlled so I can drop it back onto a fresh machine in one command. It's also packaged so anyone can install the same kit, tuned for either a developer or a non-coder.

Deep docs live with the code they describe: the workflow plugin's internals (the autonomous loop + the context watchdog) are in plugins/workflow/README.md, and the slash-command kit is in plugins/personal-tools/README.md. This file is the front door: what it is, how to install it, and how the pieces fit.

Quickstart

Both setups are user-wide — they install into ~/.claude, not a project folder, so the kit follows you across every project. Pick one and you're done; full install details are in Install (full) below.

Option A — let Claude Code do it (no terminal needed)

Open Claude Code and paste one of these:

Non-developer:

I'm not a programmer and I want to start a project with your help. Please set up Claude Code for me: read the setup instructions at https://github.com/CrazyWillBear/my-dotclaude/blob/main/AGENT_SETUP.md and follow the non-developer steps. Install everything and explain what you're doing in plain English.

Developer:

Set up Claude Code with my full kit. Read https://github.com/CrazyWillBear/my-dotclaude/blob/main/AGENT_SETUP.md and follow the developer steps: install the plugins, the Playwright MCP, and the gh allowlist, and write the global CLAUDE.md.

Option B — run a script

# developer (macOS / Linux / WSL)
curl -fsSL https://raw.githubusercontent.com/CrazyWillBear/my-dotclaude/main/setup/setup-dev.sh | bash
# non-developer
curl -fsSL https://raw.githubusercontent.com/CrazyWillBear/my-dotclaude/main/setup/setup-simple.sh | bash

macOS / Linux / WSL only. On Windows, run the scripts under WSL.

Then restart Claude Code so it loads the plugins.

What's in here

• Global CLAUDE.md (global/CLAUDE.md → ~/.claude/CLAUDE.md) — my machine-wide working rules: test-driven, small diffs, ask before anything destructive, never commit secrets. (The non-developer kit installs a plain-English CLAUDE.md instead.)
• personal-tools plugin (plugins/personal-tools/) — my own slash commands and subagents: /explain, /diagnose, /my-review, /dedup-search, /init-python-project, and the human-in-the-loop dev front-end /grill-me → /to-prd → /to-issues plus /handoff. It also ships the worktree guard — a PreToolUse hook that keeps writes out of a repo's primary checkout and into a per-task worktree (EnterWorktree), so parallel sessions never collide, plus a SessionStart GC backstop for crash-orphaned worktrees. Full reference: plugins/personal-tools/README.md.
• workflow plugin (plugins/workflow/) — three things in one plugin: an autonomous dev loop (/orchestrate) that solves GitHub issues in parallel worktrees, a single-task plan→build→review chain (/pipeline — planner/implementer/reviewer models routed to the task's complexity tier), and a context watchdog that drives deliberate, early /clear and /handoff as the window fills. Full reference: plugins/workflow/README.md.
• caveman — third-party plugin for terse output; installed alongside the above.
• agent-sdk-dev — Anthropic's official plugin for scaffolding Claude Agent SDK apps (/new-sdk-app); installed alongside the above.
• perf and security-guidance — third-party plugins from Composio's marketplace: /perf runs a multi-phase performance investigation (baseline → profile → hypothesis → optimize), and security-guidance adds an advisory hook that flags risky code (eval(, execSync(, os.system, …) before a write. (The guidance hook blocks the first such edit per session so it gets a second look; set ENABLE_SECURITY_REMINDER=0 to silence it.)
• security-sweep — third-party, read-only security-scan skill: greps the project for secrets, injection, auth/config issues, and weak deps against OWASP / LLM / Mobile top-ten patterns.
• Setup scripts (setup/) — two user-wide installers: setup-dev (developer) and setup-simple (non-developer). Both install the same kit into ~/.claude.

How to use

What you actually type day to day. One human-in-the-loop front-end and one AFK loop, with GitHub Issues as the tracker (via the gh CLI).

The issue loop