Need AI Training/Help?CloudYeti.io/meet
MarkdownMe
Prompt review guide

System Prompt Review Checklist

Review a production system prompt for scope, tool rules, refusal behavior, output contract, private-data handling, and eval coverage.

Direct answer

A production system prompt should define the assistant's role, scope, allowed and disallowed actions, tool-use rules, output contract, refusal or uncertainty behavior, private-data handling, and test cases. A review should flag vague policies, hidden contradictions, missing boundaries, and instructions that are impossible to evaluate.

Open Improve My Prompt

When to use this

  • You are shipping a support bot, agent, content workflow, or internal assistant.
  • The system prompt grew over time and now has contradictions.
  • You need a review checklist before adding the prompt to production.

Steps

  1. Paste the full system or developer prompt.
  2. Check whether the scope and target user are explicit.
  3. Identify contradictions between policy, task, and output instructions.
  4. Add clear tool-use, privacy, and failure-mode rules.
  5. Test the revised prompt against a small eval set before shipping.

Example review

Prompt
You are a support assistant. Be helpful and answer customer questions. Use tools when needed. Never make mistakes. Keep answers short.
Review focus
Missing: product scope, allowed tool calls, escalation rules, uncertainty behavior, data-handling constraints, answer format, and eval cases. Replace impossible rules like never make mistakes with concrete behavior such as cite the source record, ask a clarifying question, or escalate.

Common mistakes

  • Do not rely on be helpful as a production policy.
  • Do not say use tools when needed without defining when and which tools.
  • Do not ship a system prompt without representative test cases.

FAQ

What should a system prompt include?
It should include role, scope, tool rules, output contract, boundaries, privacy constraints, uncertainty behavior, and examples or eval cases for important workflows.
How do I find contradictions in a system prompt?
Look for rules that cannot both be true, such as always answer and ask clarifying questions, or be brief and include every detail.
Should production prompts include evals?
Yes. Even a small set of representative inputs catches prompt regressions better than manual reading alone.

Related prompt guides