How to compile, run, and test — no guessing, no invented scripts
Domain 7 · 7% of the exam
An apprentice
at every bench.
The final domain is the closest to home: your own team's workshop. Configuring Claude Code and its kin so every engineer works with a tireless apprentice — one that reads the shop manual, follows the house style, and never touches the master switch alone.
Objectives
What you’ll be able to do
- Configure Claude tools and environments for teams
- Improve developer workflows with AI-assisted tooling
- Support debugging and operational issue resolution
The story
The shop manual
When Meera's team first turned on Claude Code, the results were chaos with excellent grammar. The apprentice was brilliant and lost: it invented build commands, ignored the house test conventions, and once formatted an entire package in a style nobody used. Theo wanted to blame the tool. Meera pointed at the wall where, in any real workshop, the shop manual hangs. "You hired an apprentice and told it nothing. What did you expect?"
They wrote the manual — a CLAUDE.md at the repo root: how to build, how to run the tests, the style rules, the architecture map, the things you never touch without asking. Shared with the whole team, versioned like code. The apprentice read it every morning and stopped guessing. Onboarding time for humans dropped too — the manual was the documentation nobody had written for each other.
Then came the night the payment queue backed up. The on-call engineer fed the logs and traces to the apprentice, which found the pattern in minutes — a retry storm from a config change. It proposed a fix. And here the workshop rule held: the apprentice diagnoses; the engineer decides. The fix was reproduced, tested, reviewed, and shipped like any other change. Fast is good. Governed and fast is the job.
Metaphor map
Images that stick
| Concept | Image | Why it sticks |
|---|---|---|
| Claude Code | An apprentice at every bench | Tireless, fast, and exactly as good as the workshop that trains it. |
| CLAUDE.md | The shop manual on the wall | Commands, conventions, boundaries — written once, read every morning, versioned like code. |
| Team configuration | Jigs and fixtures | Shared setups that make the right cut the easy cut, for every engineer identically. |
| AI debugging | The stethoscope | Hears patterns in the noise no human ear catches at 3 a.m. — but doesn’t write the prescription. |
| Permissions | The machine guard | The saw works fine with the guard on. Destructive actions get confirmation; scope stays per-repo. |
| The review gate | The master’s bench | Everything the apprentice makes crosses it before leaving the shop. The bar is the bar. |
The manual
What goes in the team's CLAUDE.md?
Toggle what to include in the shared project configuration. Some of these belong. One of them never does.
Naming, formatting, test patterns the team actually uses
Modules, boundaries, where things live and why
Generated files, migration history, vendored code — ask first
What a "claim", "adjuster", or "policy binder" means here
Database passwords and tokens "so the tool can run everything"
Workflow amplifier
Where the apprentice helps — and where the human stays
Pick a stage of the developer workflow.
Scaffolding
- AI does
- Generates boilerplate, project setup, config plumbing, and first-draft implementations from the conventions in CLAUDE.md — the blank-page problem disappears.
- Human owns
- The shape of the thing: module boundaries, interfaces, naming the concepts. Structure is a decision, and decisions have owners.
Refactoring
- AI does
- Executes mechanical transformations across dozens of files — renames, pattern migrations, dependency swaps — consistently and without fatigue.
- Human owns
- Choosing what to refactor and why, defining the target pattern, and judging when the churn is worth the risk.
Testing
- AI does
- Drafts test suites, generates edge cases humans forget, fills coverage gaps, and keeps fixtures consistent with the codebase.
- Human owns
- Deciding what correct means. A generated test that enshrines a bug is worse than no test — assertions get human eyes.
Code review
- AI does
- First-pass review: style violations, obvious bugs, missed error handling, security smells — before a human ever looks.
- Human owns
- The judgment layer: is this the right change, does it fit the architecture, should it exist at all? Approval carries a name.
Incident debugging
- AI does
- Ingests logs, traces, and diffs at inhuman speed; correlates the timeline; proposes hypotheses ranked by evidence.
- Human owns
- Owning the fix: reproduce, test, review, ship through the normal pipeline. The apprentice diagnoses; the engineer decides.
In production
Rolling Claude Code out to 200 engineers
In production · developer platform
Rolling Claude Code out to 200 engineers
- CONFIGURE
- A CLAUDE.md in every repo: build/test commands, conventions, architecture notes. Written once, versioned, reviewed like code.
- GOVERN
- Tool permissions scoped per repo; destructive commands require confirmation; no credentials in any shared config.
- ENABLE
- Internal playbooks for the big wins: test generation, refactoring, migration scripts, incident log analysis. Champions per team.
- HOLD THE LINE
- AI-written code walks the same road as human code: tests, review, CI gates. Productivity went up; the bar never moved.
Flashcards
Recall drill
Select a card to reveal the answer.
Quiz
Exam drill
Question 1
A team adopts Claude Code and complains it "invents build commands and ignores our test conventions." What is the highest-leverage first fix?
The tool lacks project knowledge, not intelligence: shared, versioned project configuration gives every session the shop manual — commands, conventions, boundaries. Model tier (A) doesn’t add missing facts; per-session prompts (B) don’t scale or stay consistent; read-only (D) removes the value instead of the problem.
Question 2
To "let the assistant do everything," a developer proposes committing database credentials into the shared project configuration. What’s the correct call?
Anything readable by the tool can end up in context, logs, or output — shared config must never hold secrets. Private repos leak (A), base64 is not encryption (C), and staging credentials are still credentials (D).
Question 3
During a production incident, the AI assistant analyzes the logs and proposes a config fix that "looks right" to the exhausted on-call engineer at 3 a.m. What should happen next?
AI acceleration compresses diagnosis, not governance: the fix ships through the same tested, reviewed pipeline as any change — that’s what keeps 3 a.m. fixes from becoming 4 a.m. outages. Blind application (A) and self-review (B) remove the human gate exactly when judgment is weakest; (C) throws away the speed you just gained.
Revision