Judging a Good Prompt
Judging a Good Prompt
How to tell if AI output is actually good — before you try to improve it.
Why Judge a Prompt?
Before you can improve a prompt, you need a way to measure how good its output is. Without clear criteria, prompting becomes guesswork — you tweak randomly and hope for the best.
Judging turns that guesswork into a repeatable skill. By scoring output against defined quality dimensions, you can see exactly what's weak, decide what to fix, and prove that one prompt is genuinely better than another.
A Simple Analogy
Think of a teacher grading an essay. They don't just say "good" or "bad" — they use a rubric: ideas, structure, grammar, relevance. That rubric makes grading fair and consistent. Judging a prompt works exactly the same way.
The 5 Quality Dimensions
Evaluate every output against these five questions. Each is a lens on a different kind of failure.
Accuracy
Is the information correct and verifiable?
Check facts, numbers, names, and claims against a reliable source.
Relevance
Does it answer what was actually asked?
A beautiful answer to the wrong question still fails.
Completeness
Are all parts of the request covered?
Multi-part questions need every part answered.
Format
Is the structure exactly what was requested?
The right content in the wrong shape breaks downstream use.
Tone
Is the voice right for the audience?
Formal, friendly, technical — the tone must fit the reader.
Use a Simple Scoring Rubric
Score each dimension from 1 (poor) to 5 (excellent), then compare prompts side by side to pick the winner. Here's a ready-to-use rubric:
| Dimension | What to Ask | Score (1–5) |
|---|---|---|
| Accuracy | Is every fact correct and verifiable? | ___ / 5 |
| Relevance | Does it address the actual question? | ___ / 5 |
| Completeness | Are all parts covered? | ___ / 5 |
| Format | Does the structure match the request? | ___ / 5 |
| Tone | Is the voice right for the audience? | ___ / 5 |
Prompt Problem or Model Problem?
When output scores low, diagnose the cause before rewriting. Is it your prompt, or a genuine limit of the model?
Prompt Issue (you can fix)
- Unclear or vague instructions
- Missing context or details
- No format specified
- Too many tasks in one prompt
Model Limit (rethink approach)
- Needs real-time data it can't access
- Requires tools the model doesn't have
- Highly specialised or niche domain
- Task is genuinely too complex
Example: Scoring Two Prompts
Ask the same task two ways, score each output, and the winner becomes obvious:
Prompt A: "Summarise this article."
Scores -> Accuracy 4, Relevance 3, Completeness 2, Format 2, Tone 3 = 14/25
Prompt B: "Summarise this article in 5 bullet points for a busy
manager. Keep each point under 15 words, neutral tone."
Scores -> Accuracy 5, Relevance 5, Completeness 4, Format 5, Tone 5 = 24/25
Winner: Prompt B (+10)How to Judge Effectively
- Judge before you improve — always measure first
- Score every dimension, not just an overall gut feeling
- Compare prompts side by side, don't rely on memory
- Keep a printable rubric handy for consistency
- Diagnose whether the fault is the prompt or the model
Key Takeaway
Judging turns prompting into a measurable skill. Score output across Accuracy, Relevance, Completeness, Format, and Tone, decide if the fault is the prompt or the model, and always measure before you improve.