Anthropic CCDV-F : Claude Certified Developer-Foundations

A Claude application that worked well in testing is now occasionally returning outputs that mention information not present in the input. The development team initially assumed the model was hallucinating, so they asked you to troubleshoot.
What would you do first?

• A. Add a system prompt instruction telling the model not to invent information, on the grounds that prompt-level instructions are the fastest fix for hallucination concerns.
• B. Replace the current model with a larger one to reduce the chance of hallucination, on the grounds that larger models tend to hallucinate less in typical applications.
• C. Examine production traces to identify whether the issue is hallucination by the model, context loss, prompt injection, or another failure mode before recommending a fix.
• D. Apply a retrieval-augmented generation pattern to ground the responses in source content before any further investigation of the production traces.

Your Claude application uses tool calling to fetch patient data and generate summary reports. The flow occasionally fails because the model returns a tool_use block that references arguments not present in the schema, and your application code does not handle this case gracefully.
How would you address this?

• A. Validate the tool_use block's arguments against the tool schema before dispatching the tool and handle invalid arguments as a recognized error path.
• B. Log invalid tool_use blocks when they occur and allow the tool dispatch to proceed, relying on the tool's own error handling to surface failures back to the application.
• C. Retry the same request repeatedly until the model returns a valid tool_use block that matches the schema as expected.
• D. Stop using tool calling entirely and replace tools with prompted text generation that asks the model to describe what it would do.

You are explaining to a stakeholder why running the same Claude prompt twice can produce slightly different results. The stakeholder is concerned this means the application is broken.
How would you address the stakeholder's concern?

• A. Tell the stakeholder the variation is caused by Claude being updated continuously by Anthropic, and that switching to a fixed model snapshot will eliminate the variation entirely.
• B. Tell the stakeholder the variation is a bug that the team will fix in the next release of the application, then create a work ticket to fix the bug.
• C. Explain that LLMs are non-deterministic by default due to sampling, and describe how the application handles this through validation, retries, or temperature adjustment.
• D. Tell the stakeholder the variation comes from network latency and that switching to a faster network connection will produce more consistent results across runs.

Your Claude application's prompt was written months ago and has not been updated. The team has discovered through evals that the prompt produces good results on common cases but underperforms on a specific category of inputs that has grown in volume.
How would you respond?

• A. Iterate on the prompt to address the underperforming category, validate the change with evals, and continue refining as needed.
• B. Tell users to avoid the underperforming category by adding warnings in the application's user interface about handled inputs.
• C. Add the underperforming category to a separate Claude application with its own prompt so the original prompt does not change.
• D. Replace the prompt with a new one aligned to the underperforming category, treating any common-case performance change as a known tradeoff.

Your Claude application has been running for several conversation turns, and you notice the model occasionally references information that was discussed many turns ago but is no longer relevant. You suspect context drift is causing the model to weight stale content too heavily.
How would you address the drift?

• A. Reset the conversation after every turn so the model loses all prior turns when generating a response.
• B. Apply compaction to summarize older portions of the conversation so the gist remains while the specifics carry less weight.
• C. Truncate the conversation so the model sees only the most recent turn during each subsequent response.
• D. Increase the context window size so all turns of the conversation remain visible to the model in full detail.