Key points
- Tool use with a JSON schema is the most reliable path to schema-compliant output and eliminates syntax errors.
- tool_choice auto lets Claude reply in text; any forces some tool call; a forced choice names the tool that must be called.
- Strict schemas prevent syntax errors but not semantic errors such as totals that do not add up or values in the wrong field.
- Required versus optional fields: a required field pushes the model to fabricate when the source lacks the value.
- Enums with "other" plus a detail string extend gracefully; an "unclear" value handles ambiguous cases.
- Format normalization rules belong in the prompt alongside the schema when source formatting varies.
Skills the exam expects
- Define an extraction tool and read structured data from the tool_use response.
- Use tool_choice any when several extraction schemas exist and the document type is unknown.
- Force a specific tool so a particular extraction runs before enrichment steps.
- Make fields optional or nullable when documents may not contain them.
- Add "unclear" and "other" plus detail values for extensible categorization.
Common misconceptions
- "Asking for JSON in the prompt is just as reliable." Prompted formatting is probabilistic; tool use is enforced.
- "tool_choice auto guarantees a tool call." Only any or a forced choice does.
- "If the schema validates, the data is correct." The schema checks shape, not meaning.
Watch for
- Parsing free-text replies as JSON and patching failures with cleanup heuristics.
- Required fields producing plausible but invented values.
- Forcing new categories into the closest enum value instead of "other".
Practice task 4.3
57 practice items cover this task in CCA Prep Studio: flashcards, matching, scenario questions and explain-it-back exercises, each tracked toward your mastery score.