claude-code-kit

A working configuration for Claude Code — skills, hooks, slash commands, agents, settings — that turns it from a chat-style assistant into something closer to an autonomous coding system. I run my entire workflow on it. This page is a tour of what's in the kit and why.

The premise

Claude Code out of the box is excellent at the loop of "ask, code, run, iterate." It's much less good at the work around that loop: planning multi-day projects, persisting context across sessions, doing structured PR review, keeping a TODO list and a PROGRESS.md in sync, knowing when to stop and ask versus when to just press on.

Most of those gaps can be filled with the primitives the SDK already gives you — Skill (a domain-specific micro-instruction set), Hook (a shell command that fires on lifecycle events), slash commands, sub-agents. The kit is just my opinionated take on how to wire them together.

What's in it

Slash commands. The ones I use every day:

• /loop — runs goal-driven iterations: a supervisor plans tasks, workers execute via batch-tasks, repeat until a stop condition is hit. Useful for "improve the test coverage of this module to 80%" — open-ended but checkable.
• /handoff — saves session state (open questions, what was just tried, what the next step is) into a markdown file, so the next session or a parallel agent can pick up exactly where you left off.
• /pickup — the inverse of /handoff. Reads the last handoff, refreshes itself, asks one clarifying question if needed, then continues.
• /review-pr — pulls a PR diff, runs a structured review (correctness / security / readability / tests), posts the review as a single GitHub comment.
• /sync — at end-of-day, walks the diff and updates TODO.md + PROGRESS.md so tomorrow doesn't start cold.
• /incident — flips into incident-response mode: diagnose first, write postmortem, then add follow-up items.

Skills. Smaller than slash commands. They're domain knowledge that activates only when relevant — claude-api triggers on Anthropic SDK code, tdd-workflow on test-first feature work, security-review on auth/payments code, e2e-testing on Playwright. The point is they don't bloat the system prompt; they get loaded on demand.

Hooks. Stop-hook checks that work isn't half-finished before the agent exits. UserPromptSubmit hook scans corrections and writes them into ~/.claude/corrections/rules.md — a learning file that compounds over weeks. There's a PostToolUse hook that runs typecheck after edits.

Sub-agents. Explore for codebase searches, code-reviewer, test-runner, type-checker, verify-app. Each one is a focused tool used by other agents — keeps the main context clean.

How I actually use it

The loop is roughly:

1. /pickup (resume yesterday's session)
2. Work the next TODO item — usually with Plan first, then implementation
3. /review-pr before committing
4. /sync at end-of-day
5. /handoff if I'm about to context-switch

For longer arcs (multi-week projects), /loop runs autonomously in batches between sessions; I check in once a day, redirect, and keep going.

Why it's public

I use Claude Code more than any other tool I own. People ask me how it's set up. Now I can say: "see the repo." It's also a hedge — if my laptop dies, three commands rebuild the entire workflow on a fresh machine.

Clone

git clone https://github.com/Waaangjl/claude-code-kit ~/.claude

(Yes, it just clones into your ~/.claude directly. Read the README first if you have an existing config you care about.)

Repository: github.com/Waaangjl/claude-code-kit ↗