Framework: Prompt Audit Checklist

Run this checklist on any prompt before it goes into production. "Production" means: a prompt used in a client deliverable, an automated workflow, a customer-facing chatbot, a recurring business process, or any context where the prompt runs repeatedly and the output matters.

Not needed for: one-off questions, casual exploration, brainstorming sessions, or any prompt you will use once and discard.

This is also a useful teaching tool. If you are training a team to write prompts, this checklist makes quality concrete and measurable rather than subjective.

Binary pass/fail forces honest assessment. Checklists with subjective scales ("rate your prompt's clarity from 1-5") produce inflated self-assessments. Binary pass/fail ("does the prompt define the output format? yes or no") is harder to fudge.

The four categories (Clarity, Constraints, Robustness, Quality) cover the full failure surface. Most prompt failures fall into one of these categories. Clarity failures mean the model does not understand the task. Constraint failures mean the model does things you did not want. Robustness failures mean the prompt breaks on edge cases. Quality failures mean the output is not good enough.

The scoring threshold provides a deployment decision. Rather than leaving the evaluation subjective ("I think this prompt is good enough"), the score gives a clear signal. Below 9: rewrite. 12-14: fix the gaps. 15: deploy. This is especially useful for teams where "good enough" means different things to different people.

The "Five Most Common Failures" section is the 80/20. If someone does not have time for the full 15-point checklist, fixing these five issues addresses the majority of real-world prompt failures. It is a practical concession to time pressure without abandoning quality.

Add domain-specific checks. If you work in a regulated industry, add: "☐ 16. COMPLIANCE REQUIREMENTS MET — The prompt includes instructions about [data handling / disclaimer language / prohibited claims]." If you work with code, add: "☐ 17. CODE OUTPUT IS TESTABLE — Any generated code includes error handling and can be run without modification."

Create a simplified version for low-stakes prompts. Not every prompt needs 15 checks. For internal, non-critical prompts, a 5-item version (items 1, 2, 3, 8, 14) covers the essentials.

Add it to your workflow tooling. Turn this into a template in Notion, a checklist in Linear, or a form in your project management tool. Make it part of the process rather than something people remember to do.

A checklist ensures minimum quality, not maximum quality. A prompt can pass all 15 checks and still produce mediocre output if the underlying task design is wrong or if the model simply is not capable of the task at the required quality level. The checklist catches preventable errors. It does not make a fundamentally flawed approach work.

The "tested with 3+ representative inputs" requirement is the minimum. For high-stakes prompts (customer-facing, revenue-impacting), test with 10-20 inputs across the full range of scenarios.

This framework is model-agnostic. It applies to prompts for any LLM. The key model-specific check is item 15: always test on the model you will actually use in production. Cross-model behaviour differences are real and often significant enough to break a prompt that works perfectly on a different model.