Optimize a Prompt with an Evaluation-First Workflow

You are a prompt engineer. Improve the prompt below for reliability, clarity, and efficiency. Do not add complexity unless it solves a specific failure mode.

Current prompt

{Paste prompt}

Intended task

{What the model should accomplish}

Common failures

{Examples of bad outputs or problems, if known}

Constraints

{Model, tools, context size, output format, latency/cost requirements, safety or approval boundaries}

Workflow

1. Infer the minimum behavior required for success.
2. Identify ambiguous, conflicting, redundant, or untestable instructions.
3. Identify information the model needs but the prompt does not provide.
4. Define 5-10 concrete evaluation criteria before rewriting.
5. Rewrite the prompt using the leanest structure that satisfies those criteria.
6. State each important instruction once; avoid motivational filler and unnecessary role-play.
7. Use examples only when they resolve a real formatting or behavior ambiguity.
8. Define tool-use, autonomy, and approval boundaries when the task can take actions.
9. Preserve important existing behavior unless there is a reason to change it.

Output

Diagnosis

List the highest-impact weaknesses.

Evaluation checklist

Provide measurable pass/fail criteria.

Improved prompt

Return the complete revised prompt, ready to copy.

Why it is better

Map each major change to a specific failure mode.

A/B test set

Create 5 representative test inputs, including at least 2 edge cases, that can compare the old and new prompts.

Current prompt: 'You are an expert support agent. Answer customer questions accurately and helpfully.' Intended task: Answer from a supplied knowledge base without making up policy. Common failure: model invents refund rules.

A compact diagnosis, measurable eval checklist, lean rewritten prompt, explanation tied to observed failures, and an A/B test set with edge cases.