Receipts. Always receipts.
Agent Cookbooks is a practitioner blog about Claude Code, MCP servers, hooks, subagents, skills, and CLAUDE.md patterns. Every post ships with the JSON, the timings, and the failure modes from real use — not summarized documentation. The bar: no post ships without at least one real artifact (a config that ran, a number that was measured, a failure that was hit).
If a post can't ship a receipt, it doesn't ship.
Who this is for
- Developers using Claude Code (and the rest of the AI-agent stack) for real work, on real deadlines, with a real bill at the end of the month.
- AI agent builders shipping production workflows — not demos.
- Practitioners who'd rather see the diff than read another five-tips listicle.
What's covered
- Claude Code: hooks, subagents, skills, settings.json patterns, CLAUDE.md, AGENTS.md, slash commands, the Agent SDK.
- MCP (Model Context Protocol) servers: setup, debugging, stdin/stdout patterns, real bills, failure modes.
- Comparisons: Claude Code vs Cursor, Aider, Codex CLI, Cline — same task, three or four ways, with timings.
- Programmatic recipes: templates, hooks, agent configs, and CLAUDE.md patterns you can copy and run today.
- Cost analyses: what real Claude Code usage looks like on a 30-day bill, with the breakdown by task type.
Why this exists
Most content about AI tooling is either polished marketing or five-tips listicles. Both leave practitioners guessing about what actually happens when you point an agent at real code, on a real deadline, with a real bill at the end of the month. Agent Cookbooks is the gap between the documentation and the actual experience.
How the skills wiki is sourced
The skills wiki documents Claude Code skills published by other authors under permissive (MIT) licenses. Each entry links the upstream repository and the specific skill folder on first reference, quotes sparingly, and adds firsthand observation rather than paraphrasing the upstream docs. The license note is preserved on every page.
What the receipts badges mean
Every skill page carries a receipts badge that states, plainly, how much firsthand use backs it:
- firsthand — written from real session notes: actual commands, measured numbers, and the failure modes hit along the way.
- generic — a plausible summary of what the skill typically produces, written without a firsthand run. No fabricated metrics.
- TODO — receipts are pending a real session. The page says so on its face rather than dressing up filler as experience.
Blog posts are held to the stricter bar — every post is firsthand. Skill entries are converted from generic or pending to firsthand as each skill gets used in real work.
Corrections & updates
Every post and skill page shows its publish date, and an updated date once it's been revised. Found something wrong or out of date?
Contact
[email protected] for corrections, suggestions, and "I'd love to see a recipe for X."