Original, first-party data on how people actually prompt AI — drawn from 971 real prompts submitted to SurePrompts (March–June 2026), scored on an 8-dimension quality rubric. Every study is openly licensed (CC BY 4.0) with a downloadable dataset. Free to cite.
June 17, 2026
A dimension-by-dimension autopsy of where the missing points go. Three fixes — structure, a role, and length — recover over half, while the most-skipped technique (examples, 97%) turns out to be the least valuable.
June 17, 2026
Split by the model each prompt targeted: people write 62% better prompts for Claude (27.8/100) than for ChatGPT (17.2) — but every model's audience is still failing, and those who pick no model at all score lowest of all.
June 2, 2026
We scored 971 real prompts on 8 dimensions. The average scored 20.5/100, 9 in 10 fell below 50, and 90% never assign the AI a role — yet rebuilding the same prompts with the SurePrompts generator lifted quality to 77.2/100, a 276% jump.
All studies are released under CC BY 4.0. Quote any figure with attribution — for example:
Source: SurePrompts, State of AI Prompting 2026 (971 prompts scored). sureprompts.com/research
Each prompt is scored 0–100 across 8 weighted dimensions by the same deterministic engine behind the free Prompt Quality Score tool. Aggregate-only — no prompt text is stored or published. The sample skews toward people already trying to improve their prompts, so the figures are likely an upper bound.
Writing about prompt quality and want a custom cut or a quote? Get in touch.
Yes. All 3 studies are published under a Creative Commons CC BY 4.0 license with downloadable datasets — quote any figure or chart with attribution to "SurePrompts, State of AI Prompting 2026." We're also happy to provide custom cuts or a quote on request.
Each prompt is scored 0–100 across 8 weighted dimensions by the same deterministic engine behind the free Prompt Quality Score tool. Aggregate-only — no prompt text is stored or published. The sample skews toward people already trying to improve their prompts, so the figures are likely an upper bound.
Yes. Each study links to an aggregate-only JSON dataset (no prompt text, no PII) you can download and analyze. They are openly licensed for reuse with attribution.
The studies are based on 971 real prompts submitted to SurePrompts (March–June 2026). Because the sample comes from people already using a prompt tool, it skews toward those trying to improve — so the quality figures are likely an upper bound; the broader average is probably lower.