Diagnose a prompt that is nearly working, and get a version two you can compare
Paste a prompt that gives inconsistent answers. Get a scored diagnosis, the changes with reasons, a rewritten version and the way to prove it is better.
- Works in
- Claude
- You need
- The prompt you are unhappy with · One output it produced · What was wrong with that output
- Written for
- chatgpt prompt optimizer
Scored by our own engine
This page, run through the audit we sell. Measured 11 August 2026.
A prompt that fails outright gets fixed quickly. The expensive kind is the one that works most of the time, because there is nothing obvious to change and the wrong instinct is to keep adding rules. This prompt scores yours on seven dimensions, names the single cause of the output you did not like, and returns a version two with the changes justified one by one.
Score first, rewrite second
The order matters more than it looks. Ask an assistant to improve a prompt and it will rewrite it into something that reads better, because that is what improve means to a language model. You will get a cleaner prompt with the same defect.
Scoring first forces the diagnosis to exist before the rewrite does. It also makes the review arguable: if you disagree that the output format is a 2, you can say so and rerun the rewrite with that constraint, which is not something you can do with an opaque improved version.
The seven dimensions are the ones that predict real failures. Six of them are about what the prompt says. The seventh, length discipline, is about what it says that does nothing, and it is the one that has to be enforced against the model’s own instincts.
The rule that keeps prompts small
Part 3 forbids any change that does not come with a behaviour it changes.
This exists because prompts rot in a specific way. Something goes wrong, a line is added. Something else goes wrong, another line. Six weeks later the instruction is 900 words, half of it defending against failures that happened once, and the output quality has quietly dropped because the model is now allocating attention across twenty rules instead of doing the job.
Requiring a named behaviour per change makes that visible while it is happening. A change that cannot name one is ceremony, and it should be dropped even when it sounds sensible.
When the answer is that the prompt is not the problem
Part 2 is permitted to conclude that no wording will fix it, and that outcome is worth more than a rewrite.
The clearest case is a prompt asking for information the model does not have: current search volumes, a competitor’s traffic, what is on a page it cannot open, anything that happened after training. The model will answer anyway. It will produce figures with the right shape and the wrong values, and a rewritten prompt produces them just as confidently.
The fix there is an input, not an instruction. Paste the data in, connect a tool that can fetch it, or run the check with something that actually measures. For the SEO and AI visibility jobs in that category, our free tools fetch the live page and report what is really there, which is the input the prompt was missing.
After the rewrite
Compare on one input, in two separate conversations. A model that has already seen version one in the same thread will smooth over the difference, which makes a bad rewrite look fine.
If you are optimising a prompt that produces published content rather than internal output, the last check is on the page rather than on the prompt. Run the result through the AI content readiness checker and see whether a clean passage can be extracted from what came out, because a prompt that scores well and produces unquotable pages has been optimised against the wrong target.
You are reviewing a prompt that is producing unreliable results. I will give
you the prompt, one real output it produced, and what was wrong with that
output. Diagnose it, then rewrite it.
Work in this order and show all four parts.
PART 1. DIAGNOSIS
Score the prompt out of 5 on each of these, and give a one line reason for
each score. Do not pad. A 5 gets one line too.
- Deliverable named: is it unambiguous what should exist when this finishes?
- Output format specified: shape, order, length, and column names if a table.
- Inputs marked: are the placeholders obvious and is it clear what goes in?
- Steps ordered: are the instructions sequenced where sequence matters, and
not sequenced where it does not?
- Failure handling: does it say what to do when the input is thin, missing
or contradictory?
- Fabrication guard: does it forbid invented figures, sources and examples?
- Length discipline: is there anything in it that does not change the
output? Quote the worst offending phrase.
PART 2. THE ACTUAL CAUSE
Look at the output I gave you and name which one of those scores explains
what went wrong. One cause, not a list. If the output is wrong for a reason
that is not in the prompt at all, for example the model could not know the
answer, say that instead and say it plainly.
PART 3. THE CHANGES
A numbered list. For each change: what you removed or added, quoted, and the
specific behaviour it should change. No change goes in this list without a
behaviour attached to it. If you cannot name the behaviour, do not make the
change.
PART 4. VERSION TWO
The rewritten prompt in a code block, complete and runnable, with my
placeholders preserved exactly as I wrote them so it drops into whatever I
have already built around it.
Then one closing line: the single input I should rerun on both versions to
see the difference, and what specifically to look for in the comparison.
Constraints:
- Do not make the prompt longer unless a scored gap requires it. Report the
word count of mine and of yours and say which sections grew.
- Do not add role framing, politeness or motivational phrasing.
- Do not change what the prompt is for. If you think it is aimed at the
wrong job, say so in one sentence at the end, and rewrite the prompt it
actually is rather than the one you would have written.
- Do not invent what my output should have said. You have one example and my
description of the problem, and that is all you know.
- Do not ask me questions first. State assumptions, labelled, and continue.
THE PROMPT I AM USING:
[PASTE THE FULL PROMPT]
ONE OUTPUT IT PRODUCED:
[PASTE A REAL OUTPUT, TRIMMED IF LONG]
WHAT WAS WRONG WITH IT:
[WHAT YOU EXPECTED INSTEAD, IN A SENTENCE OR TWO]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[PASTE THE FULL PROMPT]- The whole thing, including the parts you think are fine and any system or custom instruction sitting above it. A diagnosis run on an excerpt will confidently blame something that was already handled two paragraphs up.
[PASTE A REAL OUTPUT, TRIMMED IF LONG]- One genuine bad output. Trim the middle if it is long, keep the opening and the ending, and mark where you cut. Without an output this becomes a general critique of your writing, which is a different and much less useful exercise.
[WHAT YOU EXPECTED INSTEAD, IN A SENTENCE OR TWO]- The gap between what you got and what you wanted, stated as plainly as you can. It was too long is actionable. It was not good enough is not, and it is the input that turns this into generic advice.
How to run it
- 01Collect one bad output before you start
Rerun the prompt once and keep whatever it produces, even if it is only mildly wrong. The diagnosis is anchored on the difference between that output and your expectation, so this is the input that decides whether you get a real answer or a plausible one.
- 02Read Part 2 before Part 4
The single named cause is the finding. Version two is the convenience. Teams that skip to the rewrite keep the habit that produced the original, and the same fault reappears in the next prompt they write.
- 03Rerun both versions on the same input
The closing line names the input to use. Run version one and version two on it in separate conversations so neither can see the other, and compare. A rewrite that has not been compared is a preference, not an improvement.
- 04Reject changes with no behaviour attached
Part 3 requires a behaviour for every change. If one reads as tightened the wording for clarity, drop that change and keep the rest. Prompts accumulate ceremonial edits, and every one of them costs context that the task could have used.
- 05Keep version one until version two has survived a week
Prompt regressions show up on the inputs you did not test, which are usually the messy ones. Keep the previous version where you can find it, and if the new one starts refusing things it used to handle, the diff in Part 3 tells you exactly which line to remove.
Questions people ask
What does a prompt optimizer actually change?
In practice, four things: naming the deliverable, fixing the output format, adding a rule for what to do when the input is thin, and cutting phrasing that does not change the result. Those four cover most prompts that half work. Anything beyond them is usually a sign the prompt is being asked for something the model cannot know.
Why does it want an example of a bad output?
Because a prompt read on its own always looks reasonable. The output is the evidence, and Part 2 is required to name one cause that explains it rather than listing everything that could be improved. Without an output you get a tidy critique that may have nothing to do with your actual failure.
Will the rewritten prompt be longer?
Not by default, and it is instructed to report both word counts so you can see. Length is the most common wrong fix in prompt engineering: a rule is added for every bad output until the instruction is longer than the work, and the model then splits its attention across twenty constraints instead of doing the task.
What if the prompt is fine and the model simply cannot do the job?
Part 2 is allowed to say exactly that, and it is the most valuable outcome this produces. Asking for the search volume of a keyword, the traffic of a competitor or the contents of a page it cannot fetch is not a prompt problem. No wording fixes a missing input, and the honest answer saves you a week of rewrites.
Is this different from the prompt generator?
Yes. The [prompt generator](/prompts/chatgpt-prompt-generator/) builds one from a job description when you have nothing. This one starts from a prompt you already run, one output it produced and your complaint about it, and returns a diagnosis you can argue with. Use the generator to start, this to repair.
Unsubscribe in one click. We never pass your address on.