Few-shot examples for consistent output
When and how to add few-shot examples: format demonstrations, targeted ambiguous cases with reasoning, acceptable-versus-genuine pairs, varied document structures and informal values left null, diverse examples in <example> tags, worked examples, and tool input_examples.
Apply few-shot prompting to improve output consistency and quality
~8 min read · 11 practice questions ·
What the exam expects
- Recognise few-shot examples as the most effective fix when detailed instructions alone produce inconsistent, unactionable output, and reject restating, capitalising or prefilling as alternatives.
- Write 2-4 targeted examples for ambiguous cases (tool selection, branch-level coverage gaps, resolve-or-escalate) that explain why one action beat a plausible alternative, covering both sides of the line.
- Demonstrate an exact output format (location, issue, severity, suggested fix) and pair acceptable patterns with genuine issues to cut false positives while generalising to new code.
- Use examples to fix extraction from varied document structures and to keep informal values as raw text with null fields instead of invented numbers.
- Make examples diverse, wrap them in
<example>/<examples>tags, show intermediate steps in worked examples, and know the rules for toolinput_examples.
Key terms: few-shot (multishot) prompting (putting a few input-output examples in the prompt); targeted example (an example chosen for a case Claude gets wrong, not a typical one); contrastive pair (an acceptable pattern and a genuine issue side by side, each with its reason); worked example (problem, method, intermediate results, answer); input_examples (optional tool-definition field of schema-valid example inputs).
Key facts
- Detail
- Examples are more effective than abstract instructions for consistent output
- Detail
- Exam guide: 2-4 targeted examples for ambiguous cases; docs: 3-5 for best results
- Detail
- Relevant, diverse (edge cases), structured in
<example>/<examples>tags
- Detail
- Problem, method to apply, expected answer
- Detail
- Not supported on Claude 4.6 and later models
- Detail
- Explicit permission drastically reduces false information
input_examples validity- Detail
- Each must validate against
input_schema; invalid returns 400
input_examples cost- Detail
- About 20-50 tokens simple, 100-200 complex nested
input_examples scope- Detail
- Not supported on server tools such as web search
Sources for this section (5)
- Increase output consistencyplatform.claude.com (opens in a new tab)
- Exam guideeverpath-course-content.s3-accelerate.amazonaws.com (opens in a new tab)
- Prompting best practicesplatform.claude.com (opens in a new tab)
- Reduce hallucinationsplatform.claude.com (opens in a new tab)
- Define toolsplatform.claude.com (opens in a new tab)
- Why do three near-identical examples sometimes make extraction worse?
- When are targeted few-shot examples the right fix rather than more instructions?
- Why wrap samples in
<example>tags? - What happens if an
input_examplesentry does not match the tool'sinput_schema?
The questions are free to read. Answers with their sources and self-grading are included with the pack.
Keep reading with the Claude Certified Architect - Foundations pack
- Core concepts
- Exam traps
- Worked scenario
The pack includes:
- All 30 lessons of the Claude Certified Architect - Foundations guide, with reading progress
- Interactive decision walks and step-by-step sequence steppers
- Inline practice questions and one-click topic practice (11 questions for this lesson)
- An official source cited for every claim
- An AI tutor that answers from the guide
- Unlimited timed practice exams
Practice for the Claude Certified Architect - Foundations exam
Certifyforce is an independent study resource and is not affiliated with, endorsed by or sponsored by Anthropic. Claude is a trademark of Anthropic.