箱
OPEN THE BOX · a Claude Code setup, mistakes included

Inside: a team you can brief.
On top: four checks.

HOW TO READ IT
Round = agents (models): they build, read, judge.
Square = hooks (code): they enforce.
Hat letter = model: O Opus · S Sonnet · H Haiku
click any part or panel for the detail
CLAUDE.md  loads in every session.
Rules here · facts in knowledge/ · procedures in skills/ · corrections in memory/ +
5+
agents, one craft each
3+
guard hooks: prose explains, a hook enforces
35+
habits in memory/, one per file
4+
notes in knowledge/, INDEX.md first
2+
skills, one folder each
No length cap on a review
+
breaker reports every finding, ordered by damage. Brevity is for the verdict line only.
The memory index loads whole
+
MEMORY.md: 55 lines, ~5 KB. Budget: well under 200 lines and ~20 KB. A new line pays with an old one.
A reviewer that didn't write the code
+
CodeRabbit + breaker until both green → cold review in Opus: 1 test run, mutants on a risky change → a human reads and merges.
A guard on the way to main
+
gh-guard.sh: no merge without a human approval, none through the API, no push straight to main in your org's repos. 25 tests, all pass.
×

Text version

What the picture shows, for people and for AI. Every count matches the repo.

An open box labelled claude-setup-guide: a Claude Code setup, mistakes included, the part I'd hand to a friend. In the picture, round characters are agents (models) and square robots are hooks (code). The letter on a hat is the model: O Opus, S Sonnet, H Haiku.

  1. CLAUDE.md loads in every session. Rules live there, facts in knowledge/, procedures in skills/, corrections in memory/.
  2. 5 agents, one craft each: builder, breaker, organizer, navigator and skimmer.
  3. 3 guard hooks, because prose explains and a hook enforces: gh-guard.sh, pkill-guard.py, workflow-guard.py.
  4. 35 habits in memory/, one per file; 4 notes in knowledge/, with INDEX.md read first; 2 skills: pr-review-loop and see-the-data.
  5. Check 1, no length cap on a review: the breaker reports every finding, ordered by damage. Brevity is for the verdict line only.
  6. Check 2, the memory index loads whole: MEMORY.md is 55 lines, about 5 KB, well under the budget of 200 lines and ~20 KB. A new line pays with an old one.
  7. Check 3, a reviewer that didn't write the code: a bot reviewer and the breaker loop until both are green, then a cold review in Opus runs the tests once and, on a risky change, plants 10–20 small bugs. Then a person reads it and merges.
  8. Check 4, a guard on the way to main: gh-guard.sh blocks a merge without a human approval, merging through the API, and a push straight to main in your org's repos. Its 25 tests pass.