The PromptEval Prompt Quality Report scored over 1,000 real prompts across 12 use cases. The average was 52 out of 100, and only 8% reached "good" (75+). The strongest single predictor of a good prompt is whether it defines its output format, worth 27 points on average. In 9 of 10 prompts, the weakest dimension was robustness.
This is the PromptEval Prompt Quality Report. Over 1,000 real prompts have been scored on PromptEval, submitted by real users across use cases from customer support to healthcare to code. Each was scored from 0 to 100 on four structural dimensions: clarity, specificity, structure, and robustness. Every figure below comes from that set. No prompt text is stored; the analysis is anonymous and aggregate.
Only 8% of the 1,000+ scored prompts reached "good" (75 or higher). Fewer than 1% reached "excellent."Source: PromptEval Prompt Quality Report, 2026
How the scores break down
Here is how the scores spread across the set. The bar for "good" is 75, the point where a prompt is clear, specified, and holds up under variation.
| Score range | Share of prompts |
|---|---|
| 0 to 40 (failing) | 25% |
| 41 to 60 (below par) | 31% |
| 61 to 74 (functional but mediocre) | 36% |
| 75 to 84 (good) | 8% |
| 85 to 100 (excellent) | under 1% |
Roughly 92% of prompts never reach "good," and almost none reach "excellent." This includes prompts from people who clearly know the tools. The gap is not talent. It is a few missing pieces that repeat.
What separates a good prompt from a bad one
For each structural element, we compared the average score of prompts that had it against those that did not. These are averages across the set, not a controlled experiment, so read them as correlation. But the gaps were large and consistent.
| The prompt... | Avg with | Avg without | Point gap |
|---|---|---|---|
| Defines the output format | 58 | 31 | +27 |
| Has explicit constraints (what not to do) | 63 | 41 | +22 |
| Assigns a role or persona | 57 | 42 | +15 |
| Includes at least one example | 64 | 51 | +13 |
Prompts that define their output format score 27 points higher on average than those that do not (58 vs 31). It is the single strongest predictor in the data.Source: PromptEval Prompt Quality Report, 2026
The order matters. The widest gap by a large margin is not telling the model what shape the answer should take. It was also the single most skipped element in the whole dataset. This lines up with what the major prompting guides have said for years: Anthropic and OpenAI both push defining the output format, giving the model a role, and showing examples, and few-shot prompting has been documented since the GPT-3 paper (Brown et al., 2020). What the data adds is a price tag on each one.
The universal weak spot: robustness
Across the whole set, one dimension came out weak far more often than the other three. In 9 out of 10 prompts, the lowest of the four scores was robustness: how the prompt handles bad, ambiguous, or unexpected input.
In 9 of 10 scored prompts, robustness was the weakest of the four dimensions.Source: PromptEval Prompt Quality Report, 2026
Almost nobody writes the line "if the input is missing the actual question, ask one clarifying question instead of guessing." So the prompt works in the demo and falls apart the first time a real user pastes something messy. If you want the full breakdown of this failure mode, we cover robustness and edge cases here.
Short prompts almost never win
Length is not a virtue on its own, but the data is blunt about the floor.
| Prompt length | Avg score |
|---|---|
| Under 200 characters | 27 |
| 200 to 800 | 45 |
| 800 to 2,000 | 59 |
| Over 2,000 | 67 |
The score climbs steadily with length, not because padding helps, but because you cannot fit a format spec, a couple of constraints, and an example into one sentence. A one-line prompt scored 27 on average, squarely in the failing tier.
Some fields write better prompts than others
Broken down by use case, the pattern is telling. Fields where a wrong answer carries a real cost wrote the most careful prompts.
| Use case | Avg score |
|---|---|
| Healthcare, HR, research | 56 to 60 |
| Data analysis, content, support | 53 to 55 |
| Coding, education | 51 to 52 |
| Marketing, creative | 47 |
Marketing and creative prompts scored lowest, usually because they lean on vague quality words like "engaging" or "compelling" instead of a concrete spec. The model cannot optimize against a vibe.
Before and after, with the score
Here is a weak prompt and the same request rebuilt with the levers above. The jump shows where the points came from.
Before (score 34): "Write a support reply for this customer message." No role, no format, no constraints, and no handling for a message that is not actually a support request.
After (score 78): "You are a customer support agent for a SaaS product. Given a message, reply in 2 to 3 sentences, friendly and specific, and never promise a refund or a timeline. If the message is not a support request, reply only with: 'Could you tell me what you need help with?'"
Same task. The role, the format ("2 to 3 sentences"), the constraint ("never promise a refund"), and the edge-case rule together moved it 44 points, from failing to good.
What the data means for your prompts
Taken at face value, most weak prompts move up a tier with four additions:
- Format: state the exact shape of the answer (length, structure, fields). Worth the most.
- Constraints: one or two "do not" rules that fence off the ways it can go wrong.
- Role: one line telling the model who it is and what it is for.
- Edge case: what to do when the input is bad or missing. The line nobody writes.
You do not have to guess whether your prompt has these. Paste it into the free PromptEval evaluator and you get a 0 to 100 score across all four dimensions, with the specific gaps named. Three checks a month, no card required. It is the fastest way to see which of these four you are missing before it costs you a bad output. For the method behind the score, this guide walks through evaluating prompt quality, and this one explains the four dimensions.
Frequently asked questions
What is the most common mistake in AI prompts?
Leaving the output format undefined. In our data of over 1,000 scored prompts, ones that specified the format averaged 58 out of 100 versus 31 without, the largest gap of any single element.
Why do AI prompts fail on real input but work in testing?
Because most prompts have no rule for bad or ambiguous input. In 9 of 10 prompts we scored, robustness was the weakest dimension. The prompt works on clean demo input and breaks on the first messy real one.
Do longer prompts perform better?
On average yes, but not because length itself helps. Prompts under 200 characters averaged 27; over 2,000 averaged 67. A good prompt needs room for a format spec, constraints, and an example.
How do I know if my prompt is good?
Score it against the four structural dimensions: clarity, specificity, structure, and robustness. A prompt that defines its output, sets constraints, assigns a role, and handles edge cases lands in the good range. Only about 8% of the prompts we scored got there.