Engineer Production Agent Prompts

Workflow

1. Clarify the real production task, not just the wording problem. Ask what decision or action the model must perform, what can go wrong, and how success is judged.
2. Split the prompt stack correctly: system prompt for stable company or workflow rules, developer prompt for customer-, environment-, or run-specific context, user prompt for end-user intent when one exists.
3. Tighten the core prompt structure. Give the model: role, task, step order, hard constraints, output schema, and explicit conditions for rejection, uncertainty, or escalation.
4. Prefer readable structure over prose blobs. Use headings, bullets, markdown sections, or XML-like tags when the workflow has multiple stages, branching logic, or machine-parsed outputs.
5. Add worked examples and evals from real data. If the task is subtle, preserve a small set of representative good and bad cases instead of trying to prose-describe every edge case.
6. Add an escape hatch. The prompt should let the model say it lacks information, cannot safely decide, or needs clarification instead of forcing a hallucinated answer.
7. Create a feedback loop for improvement: metaprompt the prompt with a stronger model, inspect reasoning or debug traces when available, collect model complaints and failures, then update the prompt and eval set together.
8. Return a production-oriented prompt plan, not just rewritten prose. Include what belongs in each layer, what to test next, and where model-specific tuning may still be required.

Output

• Prompt stack split for system, developer, and user layers.
• Revised prompt structure or concrete edits.
• Examples and eval plan.
• Escape hatch / debug fields.
• Model-specific notes when relevant.
• Next iteration steps tied to observed failures or missing data.

Quality Bar

• Prefer operational guidance over abstract advice.
• Keep the response grounded in the source thesis.
• Ask follow-up questions if key inputs are missing.
• Do not repeat transcript wording unless a short quote is necessary.
• Do not confuse a better prompt with a complete product solution; evals, workflow understanding, and runtime feedback still matter.
• Do not force a single giant prompt when context belongs in separate layers or runtime data.

References

• Read references/source-context.md for thesis, triggers, and non-goals.
• Read references/transcript.md for source evidence and timestamps.
• Read references/patterns.md for concrete anti-patterns, fixes, and implementation details.