Prompt Engineering — Pitfalls Checklist

Most "the AI doesn't work" reports trace to one of nine reproducible anti-patterns. Run this checklist before shipping a prompt and again whenever a working prompt starts misbehaving in production.

How to use this skill

• Before shipping — walk the nine items, fix the violations, then run your eval set.
• When debugging — start at the symptom, jump to the matching pitfall, apply the fix.
• In code review — paste this checklist into the PR template for prompt changes.

The Nine Pitfalls

1. Vagueness

Symptom: Output varies wildly call-to-call. Generic, average-y answers.

Cause: Critical context is in your head, not in the prompt. "Write something about marketing" assumes the model knows your audience, length, channel, and goal.

Fix: Specify audience, length, format, and goal explicitly. The Prompt Frame's <context>, <constraints>, and <output> slots exist for this.