Architect Foundations · 20% of the exam

prompt engineering and structured output

Domain 3 configured the files around Claude. Domain 4 is about the output that comes back: the same shape every time, the right values inside it, and a check on both before anything downstream trusts them.

The scenario this domain runs on

You are building a structured data extraction system with Claude. It reads unstructured documents, extracts the fields a downstream ledger needs, and validates each record against a JSON schema before it loads. Invoices, contracts, support emails and research reports arrive together, and nobody labels which is which. A record the system is unsure of goes to human review instead of the ledger.

The pieces you are responsible for:

extraction tools + JSON schemasextract_metadatafew-shot examplesvalidator + retry loopMessage Batches APIhuman review queuedownstream ledger
~0JSON syntax errors once extraction runs through a tool with a strict schema, except when a response is cut off at max_tokens or refused. That near-zero is the trap, because a record can be valid JSON and still be wrong. Invoice INV-7731's line items sum to 1,140.00 under a printed total of 1,410.00, and an invoice marked "payable on receipt" came back with a due date because the schema required one. Both records parse, pass the schema and load into the ledger as correct.

Each of the six sections below closes one gap between "valid" and "right", and the table after the vocabulary shows which mechanism closes which.

Scenario switchThe code-review examples move to a second scenario, the guide's Claude Code for Continuous Integration. The guide writes them about a pipeline that reviews pull requests. Sections 4.1 (which findings to report) and 4.6 (who reviews generated code) run on it throughout. In 4.2 (few-shot examples for review findings), 4.4 (tracking which pattern each dismissed finding came from) and 4.5 (the pre-merge check that cannot wait for a batch), only one flagged part does. Each is flagged again when you get there.

Start hereThe words this domain is built on

structured output
A response your code can parse into fields, which in this domain means JSON that matches a schema.
JSON schema
A JSON document describing the shape of other JSON: field names, their types, and which fields are required.
enum
A schema keyword listing the only values a field may take, such as "freight" or "other".
extraction tool
A tool definition whose input_schema is the record you want, so the input of Claude's call is your extracted data.
tool_choice
The request parameter that says whether Claude may or must call a tool, which 4.3 covers.
few-shot examples
Worked input/output pairs placed in the prompt to show the model what to do with a case.
semantic error
A value with the right type in the right place that is still wrong, such as line items that do not sum to the total.
false positive
A finding that reports a problem that is not there, or, in extraction, a value the document does not support.
Message Batches API
An asynchronous endpoint that takes many requests at once, costs 50% less, and returns results within 24 hours.
tool_use, tool_result
The response block in which Claude calls a tool, and the block your code sends back with the tool's output. Domain 1 covers both.
stop_reason, max_tokens
The response field saying why Claude stopped, and the request's output limit, which cuts off a response that reaches it.
extended thinking
A request setting that gives Claude a budget of reasoning tokens before it answers.
Claude Code
Anthropic's command-line coding agent, which claude -p runs non-interactively in the CI scenario.
subagent
A separate Claude session with its own context window.
The question underneath most of this domain: what guarantees this property of the output?
Only findings worth reporting (4.1)Categories to report and to skip, with a code example for each severity level, rather than a request to be careful.
Consistent handling of ambiguous cases (4.2)Two to four targeted few-shot examples, each with the reasoning for its answer.
Valid JSON, right types (4.3)A tool with a strict JSON schema, plus a stop_reason check for a cut-off or refused response, since prompting for JSON only makes it likely.
Values that are actually right (4.3, 4.4)Your validator, with a retry that sends the specific errors if the value is in the document, and otherwise null or a conflict flag for human review.
Results in time, at the lowest cost (4.5)The synchronous API when someone is waiting. Otherwise the Message Batches API, submitted often enough to fit the SLA.
A review that questions the code (4.6)A second independent Claude instance, with one pass per file plus an integration pass.
Most wrong answers here use one row's mechanism for another row's problem, such as a stricter schema for line items that do not add up, which only your validator can catch.

task statements

flashcards

0 / 18 known
Card 1 of 18