LLMs1 min read

Prompting That Survives Production

Move past clever one-off prompts. Structure instructions, constrain outputs, and evaluate like you mean it.

A prompt that works once in ChatGPT is not a product feature. Production prompting is about constraints, structure, and measurement.

Write for the model, not for yourself

Be explicit about:

  • Role and goal
  • Allowed tools / sources
  • Output format (JSON schema when possible)
  • What not to do

Ambiguity is free in demos and expensive in production.

Prefer structure over poetry

Task: Summarize the support ticket.
Audience: Internal ops.
Format: JSON with keys severity, summary, next_action.
Rules: Do not invent account details. If unknown, use null.

Clear beats clever.

Eval early

Keep a golden set of 20–50 real examples. Score:

  • Format validity
  • Factual grounding
  • Refusal correctness

If you cannot measure it, you cannot ship it safely.

Version your prompts

Treat prompts like code: review them, version them, and roll them back when quality drops.