Scenario 5: Claude Code for Continuous Integration
Last verified October 27, 2026. Sources: Exam Guide v1.0 (Scenario 5; task statements 3.6, 4.1, 4.2, 4.5, 4.6; samples 10 to 12) and candidate write-ups.
The scenario
"You are integrating Claude Code into your CI/CD pipeline. The system runs automated code reviews, generates test cases, and provides feedback on pull requests. You need to design prompts that provide actionable feedback and minimize false positives." Primary domains: Claude Code Configuration & Workflows and Prompt Engineering & Structured Output.
What it tests
| Task statement | What the items ask |
|---|---|
| 3.6 CI/CD integration | claude -p (or --print) for non-interactive runs; --output-format json with --json-schema for machine-parseable findings; CLAUDE.md as the way to give CI runs project context (testing standards, fixtures, review criteria); passing prior findings into re-runs so only new or unaddressed issues are reported; passing existing test files so generated tests do not duplicate scenarios; using an independent instance to review code instead of the session that wrote it |
| 4.1 Explicit criteria | Replacing "be conservative" with categorical criteria; temporarily disabling a high-false-positive category to restore developer trust; severity definitions with concrete code examples |
| 4.2 Few-shot | Examples that contrast acceptable patterns with real issues so the model generalizes; examples that fix output format |
| 4.5 Batches | Message Batches for nightly or weekly jobs (50% cheaper, up to 24 hours, no latency SLA), never for blocking pre-merge checks; custom_id for resubmitting failures; no multi-turn tool use inside a batch request; submission-interval arithmetic against an SLA |
| 4.6 Multi-pass review | Per-file passes plus a cross-file integration pass instead of one pass over 14 files; a second instance without the generator's reasoning context |
What surprises candidates
Several passers say this block carried more git and pipeline framing than the samples suggest: reviews that re-run on every push, pre-merge versus nightly timing, which commit's diff to review, and session isolation between the job that generates code and the job that reviews it. The judgment questions are about matching the API to the latency requirement and about what to put in context, not about YAML syntax.
Answer patterns the exam rewards here
- The hang in a pipeline is fixed by
-p, not by invented flags (CLAUDE_HEADLESS,--batch) or stdin redirection. - Parsing free text with regular expressions loses to
--output-format jsonplus--json-schema. - Duplicate PR comments are fixed by giving the re-run the prior findings, not by skipping later pushes or deduplicating strings.
- Low-value generated tests are fixed by documenting standards and fixtures in CLAUDE.md, not by post-processing filters.
- The same session rarely finds bugs in its own code; an independent review instance does.
Practice the whole block with the free mock, or drill task statement 3.6 directly from the targeted-drill card.