- A short prompt is often enough; larger or more important tasks should include only the parts that matter.
- Goal: Describe the result, not detailed steps unless the process itself matters.
- Context: Add only information or sources that could change the result, and explain what the model should take from each. Too much context can confuse the model.
- Output: Name the audience, format, length, and intended use.
- Boundaries: State what must stay unchanged, what to avoid, and what requires approval.
- For difficult, verifiable work, treat the prompt as both an acceptance test and a search policy.
- Define exactly what counts as done.
- List plausible but insufficient results that do not count.
- Diversify before converging: start with independent approaches, keep incompatible routes alive, and share ideas only after each route’s strengths and gaps are clear.
- Turn likely mistakes and edge cases into an adversarial checklist.
- Actively search for counterexamples and have independent reviewers challenge candidate answers.
- Mark a route as blocked when it only moves the problem into an equally hard unproved claim, reopen it only with a materially new mechanism.
- Iterate through
attempt → failure → diagnosis → new approach → draft → audit → repair.
- Require concrete evidence, reject vague optimism, and report the exact remaining gap when nothing passes the acceptance test.
- Designing prompts is an iterative process that requires experimentation.
- Start simple. Review the result and ask for the specific change you want.
- For important work, ask the model to check the result and flag missing information instead of guessing, then review it yourself.
- Learn about the advanced prompting techniques.
- English is becoming the hottest new programming language. Use it.
- Use comments to guide the model to do what you want.
- Describe the problem very clearly and effectively.
- Make the model ask you more questions to refine the ideas.
- If you want to force some “reasoning”, ask something like “is that a good suggestion?” or “propose a variety of suggestions for the problem at hand and their trade-offs”.
- “Prompt engineering” will have a large impact on the usefulness of an agent.
- Follow Prompt Engineering Guide, Brex’s Prompt Engineering Guide, and OpenAI Best Practices. Also some more on GitHub.
- Learn from leaked System Prompts.
- Sometimes, you should let the models write their own prompts.