Context Engineering - The 30-Day Claude Code Playbook

Build 11 tools that eliminate context rot, survive restarts, and keep your AI sharp.

De: J Cook

Narrado por: Virtual Voice

Hour one, Claude Code is brilliant. Hour six, it's guessing. You open a new chat and re-explain the whole project. Again.

Meanwhile, developers on Reddit are running agents that stay sharp all day. Context files that survive a reset. Handoff packets that carry a dead session's work into the next one. Reset triggers with an actual token count attached. The difference isn't talent. It's deciding what the model sees and what it never should.

In 30 days you'll build the 11 tools that stop the fade. This weekend you'll instrument your window and see your real fill percentage for the first time. By day 7 your CLAUDE.md is smaller and every rule left in it earned its place. By day 14 a dead session hands off to a fresh one without you re-explaining anything. By day 30 the whole system is on your calendar and your token bill has a before and an after.

This isn't another context engineering book. There are about 35 of them on Amazon right now. The best reviewed one has 17 ratings. One is 42 pages. One promises to cut your token costs by 40 percent and never shows the math. None of them let you check. You can't tell whether the file you just wrote helped or hurt, and neither can the author who sold it to you. Chapter 3 builds the test first, and every technique after it has to beat your own baseline or it doesn't ship.

All eleven come with a free companion GitHub repo: one folder per chapter, the scripts transcribed from the book, and a first command that reads your own session logs with no key and no account.

Here's what you'll build:

1. A context inventory that shows what is actually eating your window.
2. A repeatable test that proves whether a context change helped or hurt.
3. A live gauge with three thresholds: watch, hand off, stop.
4. A CLAUDE.md that got smaller, where every surviving rule has a number behind it.
5. A rot detector that catches your agent reasoning off its own failed attempts.
6. A written reset trigger, so you stop guessing when to start over.
7. A handoff packet the next session checks against your repo before it continues.
8. A subagent split that keeps your main window flat while the work gets done.
9. A retrieval policy that fetches the current file instead of trusting a remembered one.
10. A token budget with a measured before and after. One builder profiled three months of sessions, found a single file being read 118 times, and cut $15,000.
11. A team context bundle with freshness rules, so shared context stops going stale.

This book is 177+ pages and 11 working tools, measured on your repo with your code. Scroll up and grab your copy.