Prompt engineering as a discipline is settling. After thousands of production prompts across our AI tools, six patterns keep coming back.
- Role + task + constraints. "You are a legal editor. Rewrite this clause in plain English. Under 60 words."
- Show, don't tell. Include two or three examples of the output format you want.
- Chain of thought. Ask the model to think step by step before answering. Improves accuracy on multi-step reasoning.
- Refuse-and-ask. Tell the model to refuse to answer if inputs are missing, and to ask for what's missing instead of hallucinating.
- Output schema. For anything you'll parse programmatically, ask for JSON with an explicit schema.
- Grounding. Paste the source (paragraph, code, PDF text) into the prompt and instruct the model to use only that source.
Combine them. Almost every hard prompt in production is a combination of these six.
