In the continuous integration scenario, Claude runs inside automated pipelines: reviewing pull requests, fixing lint errors, generating release notes or processing many items overnight. Because nobody is watching each run, questions set here focus on structured output, tight permissions and reliability under failure, drawing on the Claude Code, prompt engineering and context management domains.
The setup you should picture
A pipeline step invokes Claude Code non-interactively with a prompt, asks for machine-readable output, and parses the result to post review comments or fail the build. Larger jobs, such as classifying thousands of records, run as batches on a schedule. A separate review pass may examine changes produced by an earlier step.
Decisions the exam probes
Headless invocation. Non-interactive mode with a prompt flag and JSON output is built for scripts. It should be paired with an explicit allowlist of exactly the tools and commands the job needs.
Permissions without a human. In an interactive session a developer approves risky actions. In CI nobody does, so broad shell access is denied and credentials are limited to what the job requires. Bypassing permission checks "so the job never stalls" is a classic wrong answer.
Parseable, complete output. Results are validated against a schema before use, and a "max_tokens" stop reason is treated as truncation rather than success.
Independent review. A review that runs in the same context as the code it reviews tends to share its assumptions; a separate pass that sees only the diff and the requirements is more useful.
Large jobs. Non-urgent bulk work fits the Message Batches API. Every item has its own status and can be retried idempotently; transient errors are retried with backoff and persistent failures go to a dead-letter queue.
Common traps
- Running an interactive session for every pull request.
- Disabling permission checks to avoid prompts in automation.
- Assuming JSON is complete because it starts with a bracket.
- Restarting a 10,000-item job from the beginning after one failure.
- Retrying rate-limited requests immediately in a tight loop.
How to practise
Script a headless Claude Code run that reviews a small diff and returns findings as JSON, with only read tools allowed. Make the output deliberately long enough to truncate and confirm your script notices. Then write a small batch job with per-item status and simulate a few failures to see how your retry logic behaves.
The six scenarios are listed in Anthropic's official exam guide (version 1.1), with four drawn per sitting. The emphasis described here is our own study guidance, not official exam content.