Prompt Improver

About the Prompt Improver

There is a particular frustration in having a prompt that almost works. It returns the right subject in the wrong shape, or the right shape padded with invented statistics, or an answer that was excellent once and has been mediocre every time since. Starting again from a blank box feels like giving up on the parts that were fine.

Paste the prompt here instead and it comes back rebuilt, with its intent left alone. The rewrite works through a fixed list of the things that actually go wrong with prompts, in the order they cause damage. Vagueness first: an instruction that names a topic but not an audience, a length or a purpose has left the most consequential decisions to a machine that will make them blandly. Then output shape — if you need a table, a word count or valid JSON, an answer that arrives as friendly paragraphs is a failure the prompt caused. Then missing constraints, which is where invented facts come from: a model told what to write and never told what not to invent will fill a gap rather than leave one. Then assumed context you have in your head and never typed. Then outright contradictions, of which asking for something brief and comprehensive in one breath is the most common. Finally the politeness padding, which costs tokens and changes nothing.

There is a box for what goes wrong with it now, and filling it in changes the result more than any other setting on the page. "It answers in bullet points when I want prose" is a fault that can be fixed directly and specifically; without it the rewrite has to guess at your complaint from the prompt alone.

Leave "Explain each change" on the first few times. Seeing which of your habits keep coming up is how you stop needing this page.

How to use it

1. 1Paste the prompt you are using now into the box.
2. 2Describe what goes wrong with it in the options — the more precisely, the better the rewrite.
3. 3Leave “Explain each change” on if you want to learn from the edit rather than only take it.
4. 4Run it, then read the rewritten prompt before you use it: check it did not narrow your intent.
5. 5Test both versions on the same task. The improved one should fail less often, not merely read better.

Questions

Will it change what I was asking for?

It should not, and that is the first thing to check. The instruction is to sharpen the request without altering its subject or adding requirements you plainly do not want. If the rewrite has quietly narrowed the scope — turning a general request into a specific one — say so in the box describing the problem and run it again.

My prompt is very long. Is that a problem in itself?

Length is not the fault; disorganisation is. A long prompt that repeats itself, contradicts itself in the third paragraph, or buries the actual task under context tends to produce worse answers than a short one. Expect the rewrite to be shorter, but only because it removed duplication rather than requirements.

Does it work on system prompts as well as ordinary ones?

Yes, and system prompts benefit more, because they are written once and then run thousands of times without anyone rereading them. Ambiguity that you would notice and correct in a conversation just keeps producing the same drift.

Why does it sometimes remove my example?

It should not remove a good one. Examples are among the most effective things a prompt can contain. It will remove an example that contradicts the instructions around it, since a model shown one thing and told another will usually follow the example.