live benchmark · updates continuously

The State of Prompt Quality

Every prompt evaluated on PromptEval becomes an anonymous data point: one score, 0–100, from a fixed rubric across clarity, specificity, structure, and robustness. This page is the running total. Measured, not surveyed.

1,025
prompts analyzed since May 2026
54
average score
60
median
10.5%
score 75+ (production bar)
the live picture

Four dimensions. One keeps failing.

Figures below update as new evaluations come in. For citation-stable numbers, use a quarterly edition. Those never change after publication.

Structure64.6/100
Clarity63/100
Specificity57.6/100
Robustnessweakest for 96% of prompts31.6/100
score distribution
75 · production bar
050100

34.9% of prompts score below 50. 10.5% reach the production bar.

headline findings · q3 2026 edition

What the data says

Frozen on July 11, 2026 at n = 1,018. These numbers will not change under you. Cite them freely. Full analysis in the Q3 2026 edition →

54/100

The average prompt score. Median: 60. Only 10.5% clear 75, the common bar for production use.

#average-score
96%

share the same weak spot: robustness. It averages 31.5/100, half the score of clarity (63) or structure (65). Prompts are written for the happy path.

#robustness-gap
85%

of prompts have no defense against messy input. Bad-input resilience averages 30/100; edge-case coverage, 32/100.

#messy-input
+29 pts

for declaring an output format. Prompts that say what the answer should look like average 60; prompts that don't, 31. One in five still skips it.

#output-format
5%

of prompts include an example. The most underused lever in the dataset, worth +10 points on average.

#examples
45%

of prompts set no constraints. Saying what the model must not do is worth +24 points on average.

#constraints
what moves the score · live

The cheapest points people leave on the table

Four things anyone can add to a prompt, ranked by average lift in this dataset. The bar shows how many prompts actually use each one.

+28
avg. lift
Output format

Say what the answer should look like: a list, a table, three paragraphs, JSON.

79% use it
59.8 with
31.3 without
+24
avg. lift
Constraints

Boundaries and rules: what to avoid, what never to do, hard limits.

55% use it
64.5 with
40.8 without
+17
avg. lift
Persona

Who the model is: role, expertise, point of view.

71% use it
58.7 with
42.2 without
+10
avg. lift
Examples

One worked example of the output you want. The rarest lever, and it is free points.

5% use it
63.7 with
53.3 without
breakdown · live

Who writes what

Share of classified prompts by use case. Segments under n ≈ 150 appear here but don't get standalone claims. Sample discipline is part of the methodology.

Content creation
33.4%53.5
Education
24.9%55.3
Other
8.3%50.4
Sales & marketing
6.8%53
Healthcare
4.8%58.3
Creative writing
3.8%49.5
Customer support
3.6%53
Coding
3.3%53.7
Data analysis
2.5%54.3
HR & recruiting
2.3%57.3
Research
1.9%55.5
AI agents
1.6%51.8
Legal & compliance
1.2%57.9
Finance
1.0%60.1
Productivity
0.6%37.2
shareavg. score
English 74%Arabic 19%Portuguese 6%+6 more languages68% system prompts
editions

Frozen snapshots, quarter by quarter

The hub you're reading updates continuously. Editions are frozen at quarter close and never edited. Cite an edition when you need a number that holds still.

Q3 2026 · CURRENT
The Robustness Gap

The inaugural edition. n = 1,018 · why 96% of prompts share one weak spot, and the four levers that separate 31 from 60.

Q4 2026 · UPCOMING
Publishes January 2027

First quarter-over-quarter comparison: did robustness move?

methodology

How these numbers are made

01
Where the data comes from

Every evaluation run on PromptEval contributes one anonymous row: the overall score, four dimension scores, and structural metadata (length, language, whether it has examples, constraints, a persona, an output format). The prompt text itself is never stored in the benchmark dataset.

02
How scoring works

An LLM judge scores each prompt against a fixed, versioned rubric: eight behaviors, two per dimension. Each dimension is the mean of its two behaviors; the overall score is the mean of all eight. Same judge configuration for every prompt, every time. Rubric changes ship as a new scoring version, never silently.

Clarityabsence of ambiguity · absence of conflicting instructions
Structurelogical organization · critical positioning
Specificityoutput definition · constraint definition
Robustnessedge-case coverage · bad-input resilience

All eight scored and ranked in the Q3 2026 edition. Only the scores leave the box: the rubric text, weights, and judge configuration stay internal.

03
Selection bias, declared

This is not "the average prompt on Earth." It is prompts people chose to submit to an evaluator, often because they suspected something was wrong. Scores likely skew lower than the true population. Read every figure here as "prompts submitted for evaluation," never "all prompts."

04
Classification

Use case, prompt type, and language labels come from a separate model pass over ~95% of rows. Unclassified rows count toward totals but not toward breakdowns.

05
Sample discipline

Segments below roughly n = 150 appear in tables but never get standalone headline claims or dedicated pages. When a segment crosses the threshold, it graduates.

cite this

Use these numbers

Everything on this page may be republished with attribution and a link (CC BY 4.0). For figures that must never move, cite the Q3 2026 edition.

PromptEval, "The State of Prompt Quality." Live benchmark of 1,025 evaluated prompts. https://prompt-eval.com/state-of-prompt-quality (accessed July 12, 2026).

Where does your prompt land?

Paste a prompt, get the same 0–100 score this benchmark is built on, plus exactly what to fix. Three free evaluations a month. No card.

Score my prompt free →
or see how the scoring works →