Prompt patterns that survive contact with prod

The prompt that dazzled in your notebook tends to wobble the moment real users arrive with their messy, adversarial, off-distribution inputs. The fix isn't a more elaborate prompt — it's a more disciplined one. A few patterns do most of the work.

Make the contract explicit

Tell the model exactly what shape you expect back and what to do when it can't comply. Ask for structured output, give it a clear "if you're unsure, return this" escape hatch, and validate the result before you trust it. A model that can say "I don't know" in a format your code understands is worth more than one that's confidently wrong.

Show, don't just tell

A couple of well-chosen examples outperform a paragraph of instructions almost every time — especially examples of the hard cases and the edge cases you keep getting wrong. Curate them from real failures, not imagined ones.

Separate the stable from the variable

Keep your instructions, your examples, and the user's input in clearly delimited sections. When everything is mashed into one blob, a single weird input can hijack your whole prompt. Structure makes injection harder and debugging easier — you can see exactly which part changed.

Budget for failure

Production prompts need a plan for the bad day: a retry with a tightened instruction, a fallback to a simpler path, a cap on cost and latency. Decide in advance what happens when the model returns garbage, because at scale it eventually will.

None of this is glamorous. That's the point. The prompts that survive contact with production look less like magic spells and more like the careful interfaces you'd design for any unreliable dependency — because that's exactly what a model is.