Категориальные примеры

When you give Claude a job that has a fuzzy boundary, vague instructions like "be more precise" rarely work. The model can't read your mind about where the line sits.

The reliable fix is categorical examples: a small set of 2-4 few-shot examples that show concrete cases of what to report and what to ignore. The model generalizes from these examples to new, unseen inputs — it does not merely repeat them.

This lesson shows how to build those examples for a classic architect task: deciding what is worth flagging and what is noise.

Before you even reach for examples, state an explicit criterion. Explicit rules consistently beat vague adjectives.

• Vague: "Review the code carefully."
• Explicit: "Flag a comment only when it contradicts the code it describes."

The explicit version draws a sharp line. Categorical examples then make that line unmistakable by pinning it to concrete cases on each side.

A good set has examples on both sides of the boundary, each with a one-line reason. Two REPORT cases and two IGNORE cases is enough for most ambiguities — 2-4 targeted examples is the sweet spot.

Each example carries three parts:

• the input (the thing being judged)
• the verdict (report / ignore)
• a short reason tied to your criterion

The reason is what lets the model generalize correctly instead of pattern-matching on surface features.

Here is a categorical example block for the comment-review task. Notice the symmetry: contradictions are reported; harmless mismatches are ignored.

The power of categorical examples is that the model generalizes. It does not need an example for every possible input.

From the four cases above, Claude learns the rule "flag semantic contradictions, ignore harmless mismatches" and applies it to a brand-new comment it has never seen, such as a docstring that claims a function returns a list when it returns a dict.

This is why few-shot examples are ideal for consistency, edge cases, output format, and reducing hallucination.

The most valuable examples sit right at the boundary, where the decision is genuinely hard. A near-miss IGNORE next to a near-hit REPORT teaches far more than two obvious cases.

For a research agent deciding which statistics to surface, you might contrast a stale figure with a current one — and note that dates often resolve apparent contradictions rather than one number simply being wrong.

Categorical examples pair naturally with structured output. When the verdict must be machine-readable, force a tool call so the result is valid JSON every time.

Use tool_choice of type "any" to guarantee the model calls some tool (guaranteeing structured output), or force one specific tool by name. A JSON Schema then eliminates syntax errors and enforces required fields.

A subtle trap: in your finding schema, mark a field required only if it is always present. If you require a field that may be absent — say a line_number for findings that aren't line-specific — the model will fabricate a value to satisfy the schema.

For categories that may grow, use an enum plus an "other" value and a free-text detail field. That keeps the output structured while staying extensible.

Categorical examples sharpen a judgment boundary. They do not invent missing information.

If the model returns an empty or wrong result because the needed fact is simply absent from the source, more examples won't help — and neither will retry-with-feedback. Retry fixes format, structural, and arithmetic errors, not absent data.

Distinguish a genuine empty result ("no matches found") from an access failure (the source couldn't be read). Treat them differently in your recovery logic.

Bake your report/ignore policy where the work happens. For a CI review job, put the examples in project-level CLAUDE.md (shared via VCS) so every run and every teammate inherits the same boundary.

In CI, run the review in an isolated session (less biased than the generation context) with -p for non-interactive output. Explicit criteria plus categorical examples are exactly how you minimize false positives.

The full recipe for a clean report/ignore decision:

• Write one explicit criterion for the boundary.
• Add 2-4 categorical examples split across both sides, each with a reason.
• Pin examples at the hard edges, not the obvious cases.
• Force structured output with a tool; require only ever-present fields.
• Remember: examples shape judgment, not missing data.

Do this and Claude generalizes your intent reliably across inputs it has never seen.

A review agent flags too many harmless comments as problems. You want it to report only genuine contradictions. Which change will most reliably tighten the boundary?

Key takeaways:

• Explicit criteria + categorical examples beat vague adjectives for fuzzy boundaries.
• Use 2-4 examples split across report and ignore, each with a reason; the model generalizes, it doesn't just repeat.
• Pin examples at the hard edges; remember dates often resolve apparent contradictions.
• Pair with structured output (tool_choice "any" + JSON Schema); require only ever-present fields, use enum+"other" for extensibility.
• Examples sharpen judgment, not missing data — absent facts need a different fix, and retry only repairs format/arithmetic errors.