What 100,000 Improved Prompts Reveal About How People Use ChatGPT & Claude

In early July 2026, Prompt Sloth crossed 100,000 improved prompts. Every one of those was a real person, mid-task, deciding their prompt wasn't good enough and clicking a button to fix it. That makes this dataset unusual: it's not a survey about how people say they use AI — it's a record of the exact moment they wanted better output.

The gap it addresses is well documented. In a Pew Research Center survey of US workers, 54% of regular AI chatbot users said the tools were highly helpful for doing things more quickly — but only 41% said the same about improving the quality of their work (Pew Research Center, February 2025). Speed is easy. Quality is where people get stuck, and quality starts with the prompt.

So here's what 107,929 prompt improvements — logged between October 2025 and July 2026 — actually show. All numbers below come straight from our own database; the methodology notes at the end explain exactly what was counted.

ChatGPT barely beats Claude

Across 14,740 fully-logged improvements (May 15 – July 17, 2026), here's where people improved their prompts:

In the broader market, ChatGPT's consumer lead is enormous. In our data it's four points. People who install a prompt improver are a self-selected group — they're using AI for work that has to be right — and that group has drifted noticeably toward Claude.

It also fits who these users are: consultants, agency owners, and solo operators whose AI output goes in front of clients and employers — the people we wrote about in our AI prompts for agencies guide.

The typical prompt is one sentence long

This is my favorite finding because it explains almost everything else.

The median prompt people type is 31 words — a single sentence, maybe two. After one improvement pass it's 65 words: the same request with the context, constraints, and output format the model needed all along. (Averages run higher than medians because a minority of users paste long documents; the median is the honest picture of a typical prompt.)

Research keeps showing how much this matters. A study from the University of Washington and the Allen Institute for AI found that trivial changes to prompt formatting alone can swing model accuracy by up to 76 percentage points in few-shot settings (Sclar et al., ICLR 2024). And the famous "Let's think step by step" study showed a single added phrase lifting accuracy on one math benchmark from 17.7% to 78.7% (Kojima et al., NeurIPS 2022). Wording isn't cosmetic. It's the control surface.

Nobody wants to become a prompt engineer

Prompt Sloth offers two paths: a one-click improve button, and a more deliberate flow where you can steer the rewrite. The vote isn't close: 88.2% of improvements (84,753 of 96,080) used the one-click path.

And people trust the results. On the platforms where we can reliably detect what happens next (ChatGPT, Claude, Gemini, Lovable, AI Studio, Grok), 88.5% of improved prompts that reached a clear outcome were actually sent — only 11.5% were abandoned.

The lesson for anyone building or buying AI tooling: prompt engineering is a specialist skill, but prompt quality is a mass-market need. People don't want a course; they want the fixed version of the sentence they already typed. We unpacked this pattern in its own essay: nobody wants to be a prompt engineer.

People improve marketing copy, not code

Every improvement runs in a mode. The distribution surprised me:

Marketing is the runaway winner among specialized modes — nearly one in four improvements. Coding barely registers. Developers already have Copilot and Cursor; the people reaching for a prompt improver are writing landing pages, client emails, proposals, and product descriptions, where tone and precision are the whole job.

Two smaller signals point the same direction. First, 26.4% of the 44,540 language-detected improvements weren't in English — led by Polish, Portuguese, Italian, Japanese, and Dutch. A lot of professionals work in English as a second language, and an improver that turns rough ESL drafts into polished prompts removes a real handicap. Second, usage has a distinct work-week signature: weekday volume averages 2,273 improvements per day against 1,688 on weekends, with Monday the single busiest day. This isn't a toy people play with on Sunday. It's part of how they work.

How usage grew to 100,000

Monthly improvements grew from 2,471 in October 2025 to 6,272 in March 2026, then jumped to 33,751 in April — the month our install base spiked — and have stayed well above the early-year baseline since (23,352 in May, 17,274 in June). The cumulative count crossed 100,000 in the first week of July 2026 and stood at 107,929 as of July 17.

Methodology

• Sources. Improvement events from the Prompt Sloth Chrome extension and the website demo, October 13, 2025 – July 17, 2026 (n = 107,929, of which 96,272 extension and 11,654 demo). Detailed per-prompt records (platform, lengths, mode, outcome) are available from May 15, 2026 onward (n = 14,740); platform, length, and mode stats use that window.
• Medians vs averages. Length stats report both; medians are more representative because a small share of very long pasted prompts skews the averages.
• Sent rate. Computed only on the six platforms where send-detection is reliable, and only over prompts with a definitive outcome (sent or abandoned). Prompts with an undetected outcome are excluded.
• Language. Detected on a subset of events (n = 44,540); percentages use that subset as the denominator.
• Privacy. Everything here is an aggregate SQL statistic. No individual prompts are quoted, and prompt text never leaves our database.

Try it on your own prompts

The whole dataset exists because improving a prompt takes one click. If you want to see what your own 31-word prompt looks like after that click, try the free AI prompt enhancer in your browser — no login needed — or install Prompt Sloth and use it directly inside ChatGPT, Claude, and 20+ other AI tools.