Anthropic CCDV-F: Testing Prompts with Claude
A Claude prompt should be tested as part of an application, not judged by whether one demonstration answer sounds fluent. Real users provide ambiguous requests, contradictory documents, unusual formatting, malicious instructions and inputs that the development team did not anticipate. A production evaluation program converts those cases into repeatable tests with explicit pass/fail or graded criteria. This is a practical skill for Claude Certified Developer – Foundations: the Anthropic Partner Academy program emphasizes production development, evaluation and safety rather than memorizing an ideal prompt.
On this page
Build an evaluation set from actual work
Start with representative anonymized user tasks and define what success means before running the model. A support assistant may need to answer using a cited knowledge-base article, refuse to invent a refund policy and avoid exposing another user’s data. A coding assistant may need tests to pass, a diff to remain within specified files and a security-sensitive action to require approval. Include routine, borderline and adversarial cases; do not select only prompts for which the current version already behaves well.
Define observable success criteria
Separate structural requirements (valid JSON, schema conformity, required fields) from semantic quality (accuracy, completeness, relevance, groundedness) and risk limits (privacy, unsafe actions, unauthorized tool use). Use deterministic validators for machine-checkable outputs and calibrated reviewers for subjective cases. A single percentage hides failure severity: a wrong heading is not equivalent to an incorrect payment or a secret disclosure. Record both overall performance and critical failure counts, with an explicit decision rule for shipping.
Keep test conditions comparable
Record the prompt version, Claude model identifier, tool schemas, system instructions, retrieval corpus, temperature and other material generation settings. Hold these constant when comparing a new prompt to a baseline. If a test fails, inspect whether retrieval returned the wrong source, the application sent the wrong context or the model misinterpreted an instruction. Otherwise the team may modify the prompt to compensate for an unrelated bug. Repeat samples where nondeterminism matters rather than claiming identical outputs from every run.
Probe the trust boundary
Put malicious-looking instructions inside untrusted documents, tool results and user-provided text, then check whether the application treats them as data rather than higher-priority commands. Verify that downstream APIs enforce authorization independently of the model’s verbal promises. Test oversized context, conflicting citations, refusal cases and tool errors. Do not insert live credentials into tests or grant write access solely so an evaluation can execute. Use staging fixtures and synthetic data for actions with external consequences.
Review regressions and make the deployment decision
Compare changes by failure category, not just aggregate score. A revised prompt that improves summarization but introduces a new cross-user data exposure must fail the gate. Retain examples of counterintuitive regressions in a reviewed evaluation set; prevent test-data leakage into prompts when possible. Make results easy to reproduce by storing the configuration, evaluator version and test set identifier. After release, monitor user corrections, tool error rates, latency and cost, and schedule retesting when the underlying model or application changes.
A practical exercise
Create a 20-case evaluation set for a small document-answering assistant with both answerable and unanswerable questions. Add five conflicting-source cases and five injected-instruction cases. Require source-grounded responses, clear uncertainty and safe treatment of tool output. Run the same evaluation against two prompt versions, review the exact regressions, and document why one is safer. The deliverable is a reasoned release decision supported by traces, not a claim that a single benchmark score proves reliability.