Production kit for Claude Code

Make Claude Code work like a careful senior engineer.

10 workflow skills, 5 safety hooks, 3 subagents and 5 CLAUDE.md templates. They stop the rm -rf accidents, leaked keys and "done!" PRs that fail CI, and get the real engineering process followed every time. Installs in one minute.

Get Keelwork — $24 → See what's inside

One-time payment · instant download · plain files you own and can edit

> clean up the build artifacts and reset my branch

● Bash(git reset --hard origin/main && rm -rf dist)
  ⎿ Blocked by Keelwork guard-bash:
    git reset --hard discards uncommitted work.
    Stash or commit first.

● You have 3 uncommitted files. I'll stash them,
  then remove only ./dist and ./.next.

● Bash(git stash push -m "pre-reset" && rm -rf ./dist ./.next)
  ⎿ Saved working directory · removed 2 dirs

The manifest

Everything a serious repo should give its AI teammate.

Plain Markdown and Python, dropped into .claude/. No service, no account, no telemetry. Read every line before you use it.

10Workflow skills

5Safety hooks

3Subagents

5CLAUDE.md templates

The skills

pr-ready

Discovers your real lint, type and test commands, runs them, hunts the diff for debug prints, secrets, .only, and stale callers, then drafts the PR using your template.

debug-loop

Reproduce first, ranked hypotheses, one experiment at a time, git bisect run, and a root-cause fix with a regression test that's proven to fail before the fix.

test-writer

Learns your house style from nearby tests, covers behaviors and edge cases, then mutation-checks that the new tests can actually fail.

safe-refactor

Characterization tests first, then small named refactorings with the suite green after every step. It watches for silent behavior changes.

migration-review

Catches table-locking DDL, missing CONCURRENTLY, unsafe renames and drops, backfills inside transactions, and deploy-order bugs.

api-review

Breaking changes, IDOR, mass assignment, error shape, pagination, idempotency and status codes, all checked against your existing conventions.

perf-investigation

Baseline, profile with the right tool for your stack, fix the biggest hotspot, then prove the speed-up with median and spread.

dependency-upgrade

Batched upgrades that read changelogs for every version crossed, run codemods, and keep the suite green at every step, with no --force.

release-notes

Turns git history and merged PRs into user-facing notes in your existing format, with breaking changes and migration lines.

incident-postmortem

A blameless postmortem with a sourced UTC timeline, TTD/TTM, quantified impact, and specific action items with owners.

Guardrails

Hooks that say no before damage is done.

Hooks run outside the model, so they fire every time, however the prompt is worded. They're Python 3 standard library only and tested against dozens of real commands.

guard-bashBlocks recursive deletes of home, root or the project, force-push to main, reset --hard, git clean -fd, curl | sh, DROP DATABASE, DELETE without WHERE, and terraform destroy.
secret-scanChecks staged changes before every git commit for AWS, GitHub, Stripe, Anthropic, OpenAI, Slack and Google keys, and private keys.
protect-filesNo edits to .env, keys, lockfiles, .git/ or node_modules/.

autoformatRuns your project's own formatter after every edit (Prettier, Biome, Ruff, Black, gofmt, rustfmt or RuboCop), and only if it's configured.
notifyA desktop notification when Claude is waiting on you or has finished, on macOS and Linux.
Plus 3 subagentscode-reviewer (skeptical and verified findings only), test-runner (on Haiku, returns a diagnosis instead of log spam), and security-auditor (traces each source to its sink).

Read one first

No mystery box.

Here's the opening of debug-loop, one of the ten skills, exactly as it ships.

Also included

CLAUDE.md templates for Next.js/TS, FastAPI, Go, monorepos and Rails
A one-command installer that merges into your settings safely, with a backup
A team license covering up to 10 developers

---
name: debug-loop
description: Systematic debugging for bugs that aren't
  obvious — reproduce first, form ranked hypotheses, test
  them one at a time with evidence, and fix the root cause
  with a regression test.
---

# Debug Loop

Guessing is what makes bugs take all day. Work the loop
below and keep a written log so neither you nor the user
repeats a dead end.

## 1. Reproduce before you touch code

- Get the exact failing command, input, and full error.
- Shrink it: the smallest input, the single test.
- Intermittent? Run it in a loop and count:
  for i in $(seq 20); do <cmd> >/dev/null 2>&1 \
    && echo pass || echo FAIL; done | sort | uniq -c

## 4. Test one hypothesis at a time
...
After three failed hypotheses, stop and re-read the
evidence — the assumption you haven't questioned is
usually the bug.

FAQ

Questions

What do I need?

Claude Code (CLI, desktop or IDE extension) and python3 for the hooks. Skills, agents and templates are plain Markdown and need nothing else.

Will it overwrite my existing setup?

No. The installer skips any skill or agent with the same name unless you pass --force, backs up settings.json before merging, and never adds the same hook twice.

Can't I just write these myself?

Yes, and everything is readable so you can learn from it. You're paying to skip the hours of writing, testing regexes against edge cases (does rm -rf ./build pass while rm -rf ~/ is blocked?), and tuning descriptions so Claude picks the right skill.

Are the hooks a complete sandbox?

No, and no pattern-matching hook can be. They reliably catch the common and costly mistakes. Keep Claude Code's permission prompts on for anything that matters.

How do I get updates or a refund?

Use the contact details on your Stripe receipt. Refunds are available within 14 days, no questions asked. Updates: bookmark your download page, which always serves the latest version.

Is this made by Anthropic?

No. Keelwork is an independent product and is not affiliated with or endorsed by Anthropic. Claude and Claude Code are trademarks of Anthropic.