The Robustness Gap — State of Prompt Quality, Q3 2026 | PromptEval<br>54/100<br>average score
31.5<br>robustness average, half of every other dimension
96%<br>have robustness as their weakest dimension
10.5%<br>clear 75, the production bar
01 · the shape of it<br>Most prompts are mediocre in a very specific way<br>The distribution is not a disaster story. Two thirds of prompts land between 40 and 79: functional, gets-an-answer territory. Only 12% are truly broken (below 30). The problem is the ceiling: just 10.5% reach 75, the score where a prompt is consistent enough to trust in repeated use. Almost everyone writes a prompt that works once.
050100
Score distribution, n = 1,018. Median 60 · 34.9% below 50 · highest score: 90.
02 · the gap<br>Three dimensions cluster in the 60s. One sits at 31.<br>The rubric scores four things: is the prompt clear, is it specific, is it well-structured, and is it robust . That last one asks what happens when the input isn't what you pictured. Empty message. Pasted garbage. A question in another language. A user actively trying to break it.<br>Structure64.6/100
Clarity63/100
Specificity57.6/100
Robustness31.5/100
This is not a tail effect. Robustness is the weakest of the four dimensions in 95.7% of all classified prompts, across every use case, every language, every prompt type in the dataset. 83.6% score below 50 on it. People write prompts for the happy path: the clean question, the cooperative user, the input that looks like the one they tested with.<br>The sub-scores say it plainly. The two lowest-scoring behaviors in the entire rubric are both robustness behaviors:<br>Bad-input resilience
30.3<br>Edge-case coverage
32<br>Constraint definition
52.2<br>Absence of ambiguity
58<br>Critical positioning
61.9<br>Output definition
63.3<br>Logical organization
66.8<br>Absence of conflict
67.6<br>average sub-score, n = 969 classified prompts
Why it matters: a prompt that scores 60 on clarity and 30 on robustness behaves perfectly in the demo and falls apart the first week of real use, because real use is where the messy input lives. The gap between "worked when I tried it" and "works" is, statistically, this gap.
03 · the levers<br>The gap between 31 and 60 is four sentences<br>We tagged every prompt for four structural habits and compared average scores with and without each. None of these require expertise. All of them are one to four sentences.<br>+29Output format79% use it<br>One line: "Answer as a numbered list" / "Return valid JSON" / "Three short paragraphs."<br>avg. 59.8 with · 31.1 without
+24Constraints55% use it<br>Tell the model what NOT to do: "Do not invent sources." / "Stay under 200 words."<br>avg. 64.5 with · 40.8 without
+17Persona70% use it<br>Give it a role with a point of view: "You are a pediatric nurse explaining to a worried parent."<br>avg. 58.7 with · 42.2 without
+10Examplesonly 5.2% use it<br>Paste one example of the output you want. One is enough to shift the distribution.<br>avg. 63.7 with · 53.3 without
The sleeper is examples. It has the smallest lift on paper (+10) but the most headroom by far: 94.8% of prompts don't have one. Everyone has heard "give the model examples"; almost nobody does it.
04 · the sample<br>Who writes what<br>Content creation and education dominate: together they're 58% of classified prompts. The dataset is 74% English, 19% Arabic, 6% Portuguese; 67% of prompts are system prompts, the reusable kind that runs many times and multiplies the cost of a weak spot.<br>Content creation
33.6%53.5<br>Education
24.6%55.2<br>Other
8.4%50.4<br>Sales & marketing
6.8%53<br>Healthcare
4.9%58.3<br>Creative writing
3.8%49.5<br>Customer support
3.6%53<br>Coding
3.1%53.8<br>Data analysis
2.5%54.2<br>HR & recruiting
2.3%57.3<br>shareavg. score
Segments under n = 20 (finance, legal, productivity, AI agents, research) are held back from this table. Small samples make loud, wrong claims. Language split: EN 74% · AR 19% · PT 6%.
05 · what to do with this<br>If you only fix one thing, fix the crash<br>The data suggests an order of operations for anyone improving a prompt they actually reuse:<br>Add one line for the bad day: "If the input is empty, unclear, or off-topic, say so and ask for clarification instead of guessing." That single sentence addresses the behavior 85% of prompts score under 50 on.<br>Declare the output format. Biggest measured lift in the dataset (+29), costs one sentence.<br>Set two or three constraints: what the model must never do. (+24)<br>Paste one example of a good output. You'll be in the top 5% of prompts by that habit alone. (+10)
methodology, briefly<br>Every figure comes from 1,018 prompts submitted to PromptEval's evaluator between May 20 and July 11, 2026, scored 0–100 by a fixed LLM-judge rubric across four dimensions. Classification labels (use case, type, language) cover 969 rows via a separate model pass. Prompt text is never stored in the benchmark dataset. Selection bias, declared: these are prompts people chose to submit for evaluation, often suspecting a problem, so scores likely skew lower than the general population....