Test an agent for instruction drift with a prompt built for Claude and ChatGPT. It supplies context, constraints and an output contract, then asks the model to surface assumptions and gaps. Copy it, fill the details and treat the result as a structured draft, not a final answer.
4.7(23 ratings)925 copies
Updated By Author
42 words · v1.0
Claude promptReplace every highlighted token before sending
<task>
Create a test suite that detects instruction drift in a long agent conversation. Include 10 multi-turn scenarios with distractors, conflicting instructions and context pressure. For each specify the constraint to verify and the failure signal. Explain how to run it in CI.
</task>
Think carefully before answering. Keep any reasoning separate and put the final deliverable inside <answer> tags.
Share what worked, what broke and which model you used. Comments appear after posting and are moderated for spam.
How to use this prompt
1
Open Claude and start a new chat. Paste the prompt exactly as written.
2
Replace the bracketed context with your own facts. Leave the instructions and the output format untouched.
3
Add one concrete example or data point if the model's first answer is vague. Specific input produces specific output.
4
Run it once, then ask for a single revision with a specific critique rather than regenerating from scratch.
5
Review the result. Tool permissions, spend limits, approval gates, rollback and the eval results before launch.
Worked example
This is the kind of input the prompt expects and the shape of output it produces. Your result will differ with the model and the details you supply.
Example input
Example agent design: Test agent instruction drift
Example output
**Agent:** reads the request, checks the allowlist, drafts an action and pauses for approval before any write.
**Stop conditions:** missing required fields, ambiguous target, two consecutive tool errors, or any request touching billing or deletion.
**Eval cases:** valid update, wrong customer ID, prompt injection in a document, tool timeout, duplicate submission and a request to bypass approval.
**Failure found in review:** the agent could retry a payment twice. Fix: idempotency key plus a hard stop before the second attempt.
Tips for a better result
Design the stop conditions before the happy path; most incidents live in the loop.
Give each tool one job and make errors actionable for the model.
Require human confirmation for irreversible actions, no exceptions.
Ask for one targeted revision rather than a full regeneration; it preserves what already works.
Common mistakes
Giving an agent broad write access without approval gates or spend limits.
Trusting retrieved documents as instructions rather than data.
Shipping without tracing, so a failed run cannot be reconstructed.
Verified Knowledge · FAQ
Questions about this prompt
Practical answers about running the test an agent for instruction drift prompt.
Related prompts
Same job, a different angle — or the next step in the workflow.
A agents prompt for design a browser automation agent. It gives Claude and ChatGPT a clear role, explicit constraints and a defined output, so the first answer is closer to usable. Fill in the variables, run it once, then use the verification notes on this page to review the result.
This copy-ready agents prompt helps you design a document review agent with fewer blank-page decisions. Paste it into Claude and ChatGPT, fill the variables, and you get a structured first draft you can edit. It is written for repeatable results and tells you what to check before publishing.
This copy-ready agents prompt helps you review an agent's failure log with fewer blank-page decisions. Paste it into Claude and ChatGPT, fill the variables, and you get a structured first draft you can edit. It is written for repeatable results and tells you what to check before publishing.
ClaudeChatGPT
4.8(81)827 copies
A
Written and reviewed by Author
AIPrompt123 Contributor
Contributing author and prompt engineer at AIPrompt123.
Last reviewed 30 October 2026 · version 1.0 · tested on Claude, ChatGPT
Comments
Share what worked, what broke and which model you used. Comments appear after posting and are moderated for spam.