What we found
Most people write to ChatGPT, Claude or Gemini the way they would write a sticky note: “Write my weekly status update.” The model fills every gap with a guess, which is why the first answer so often needs three more rounds. We wanted to know how big those gaps are, task by task, so we measured them.
- one-line requests score 1 out of 10
- 249 of 319
- the median length of a one-line request
- 8 words
- one-liners that say what to do when a fact is missing
- 0
- the median length of the same tasks as complete prompts
- 341 words
The one-liners came from the titles of the repository’s own prompts, such as “Write a job description” or “Summarize a long report for a busy reader”. That is close to what people type. Scored the same way, 249 got 1 out of 10, 62 got 2 and 8 got 3. The average was 1.24.
What a one-liner leaves out
The ten checks ask whether a prompt says what to produce, who it is for, the facts the model cannot guess, the length, the format, what to avoid, what good looks like, an example, what to do when something is unclear, and whether it is laid out the way the model reads best. A one-line request passes the first check about two times in three, and almost nothing else.
319 one-line requests, scored with the Prompt Grader’s free ten checks.
Two gaps stand out. Almost no one-liner gives the model the facts it needs, so it invents plausible ones. And not one of the 319 says what to do when something is missing, so the model never asks; it guesses and carries on. Those two gaps cause most of the wrong numbers and made-up details people complain about.
One prompt, five sentences
Here is one task built up a sentence at a time, with the score after each step. The scores come from the same free checks, so you can paste each version into the Prompt Grader and get the same numbers.
Scores from the Prompt Grader’s free ten checks. The full version is 87 words.
The full prompt is 87 words, about a minute of typing:
Write my weekly status update for my director, who reads it on a phone before the Monday meeting. Three bullets on what moved, then one risk with its owner and date. Under 120 words. Use only the notes below and do not invent numbers. If something is missing, ask me first. A good update lets my director see in one minute what needs a decision. Match the tone of last week's update, pasted below as an example.
Notes: [paste your notes]
Last week: [paste last week's update]The repository has a tested version of this one, the weekly status update prompt, with the blanks marked.
The two additions that matter most
Most people will not write 87 words every time, and they do not need to. We added the same two short additions to all 319 one-liners and scored them again.
Same 319 tasks, same ten checks. The complete prompts are written to pass them; see How we measured it.
The first addition says who the answer is for, how long it should be and what shape it takes: “For my team. Under 150 words, as three bullets.” That alone triples the average score. The second does the most for accuracy: “Use only the facts below. If something is missing, ask me first.” It gives the model permission to stop and ask instead of inventing, which is the one thing a one-liner never says.
Then paste the facts underneath: the notes, the email thread, the figures. A model that has the material and is told not to guess is a very different tool from one that has neither.
How to rate your own prompt
Read your prompt against these ten questions, or paste it into the free Prompt Grader, which asks them for you and builds the better version as you answer.
- Does it say exactly what to produce?
- Does it say who the result is for?
- Does it supply the facts the model cannot guess?
- Does it set the size: length, count or depth?
- Does it name the format?
- Does it say what to leave out or avoid?
- Does it say what a good result looks like?
- Does it give an example to match?
- Does it say what to do when something is unclear or missing?
- Is it laid out the way the model reads best, with the material set apart from the instructions?
A prompt that answers six of these is good enough for most everyday work. For something you will reuse every week, aim for all ten once, save it, and fill in the blanks next time. The guide to the ten checks explains each one with examples.
How we measured it
We used the 319 prompts in the repository as of October 24, 2026. For each one we scored three things with the Prompt Grader’s free ten-check estimate, read for any chat model: its title as a one-line request, the title with the two additions above, and the complete prompt.
Two limits are worth knowing. First, the free estimate looks for signs of each check in the wording; it does not run the prompt or judge the answer, so a prompt can pass a check with weak content. It is a measure of what a prompt says, not of how good the result will be. Second, the complete prompts were written to pass these checks, so their 9.98 shows what a complete prompt looks like rather than proving the checks predict quality. The one-line and two-addition figures are the useful part: they show how much a typical request leaves the model to guess.
Anyone can repeat this. The scoring code runs in your browser on the Prompt Grader, and the same checks are available to AI assistants and developers through the free MCP server as grade_prompt.
Key takeaways
- Written as one line, 249 of 319 everyday work requests score 1 out of 10, and none scores above 3.
- A one-liner almost never gives the model the facts, and never says what to do when one is missing.
- Saying who it is for, how long and what shape triples the average score.
- “Use only the facts below. If something is missing, ask me first.” does the most for accuracy.
- For prompts you reuse, write all ten checks once and save the result.