Locked In Kit: Stop Babysitting Claude Code

Claude stops when the work looks done. This kit makes it stop when the work is done.

Anthropic's own Claude Code best-practices guide puts one habit first: give Claude a check it can run. Claude Code's creator says it improves quality 2–3×. Almost nobody sets it up. This kit does it for you, then adds the guardrails and context habits that make your usage last.

THE FLAGSHIP PIECES

• verify-gate: a Stop hook. When Claude tries to finish, it runs your tests. If they fail, Claude keeps working. It has loop protection built in, and it skips the run when nothing has changed.

• /lock-in: Claude interviews you about the hard parts and writes a spec with one command that proves the feature works. It arms the gate, plans, builds step by step, then has a fresh reviewer check the result against the spec.

• /handoff → /clear → /pickup: clear the context without losing your place. Smaller context means your limit lasts longer.

EVERYTHING INSIDE (43 FILES)

• 12 skills that load when the task matches: ship-it, bug-hunt, test-writer, pr-writer, safe-refactor, security-pass, explain-codebase, landing-page, db-migration, perf-audit, context-saver, commit-craft

• 6 workflows you trigger by name: /lock-in, /handoff, /pickup, /mvp, /second-opinion, /cleanup

• 6 subagents: code-reviewer, researcher (runs on Haiku to keep your main context small), architect, debugger, test-runner, ui-polisher

• 5 hooks (Python, no dependencies): verify-gate; a block on rm -rf of broad paths, force-push to main, reset --hard, DROP and curl | sh; a secrets guard for .env, keys and credentials; auto-format; a desktop notification when Claude finishes

• 8 CLAUDE.md templates: short, and only rules Claude can't work out itself. Next.js, FastAPI, Express, Expo, Chrome extension, static site, data science, plus a global personal one.

• The Locked In Method playbook: the five-step loop, CLAUDE.md that changes behavior, context hygiene, a decision table, 10 prompts that work. Sources are linked.

TESTED. The safety hooks pass 165 test cases, including commands that must not be blocked, like rm -rf node_modules. The verify gate passes 15 scenarios, including timeouts and loop limits.

Works with any plan that includes Claude Code. Mac, Windows (with Git for Windows), Linux. Everything is plain Markdown and Python, so you can read and edit every line.

14-day refund if it doesn't save you time. Free updates as the kit grows.

Independent product. Not affiliated with or endorsed by Anthropic.