DEV Community

Miha Gutman
Miha Gutman

Posted on

I’m testing whether prompt improvement tools accidentally change what users actually mean

Most prompt improvement tools optimize the prompt.

I'm interested in a slightly different problem:

What if the improved prompt no longer represents what the user actually meant?

The problem

Real prompts are rarely perfect.

People write things like:

Find me a good laptop for work

or:

Help me answer this email

or:

Make a plan for my business

The request may be vague, incomplete, badly structured, or missing important context.

A system can easily turn that into a beautifully structured prompt.

But "more structured" does not necessarily mean "more correct."

The rewritten version may:

  • invent requirements the user never gave;
  • remove something important;
  • misunderstand the actual goal;
  • silently choose between two possible interpretations;
  • add constraints that sound reasonable but aren't true.

That is the problem I'm experimenting with.

What I'm building

I'm working on an early project called Prompt Teacher.

Instead of treating prompt improvement as simply:

raw prompt → better wording

I'm exploring something closer to:

raw request → understand intent → detect missing context or ambiguity → prepare a stronger request → preserve the original intent

The important part is the last one.

A longer or more professional-looking prompt is not a success if it changes what the person wanted.

I need prompts I didn't invent

Testing this myself has an obvious weakness.

I already know what I meant when I write the test prompt.

So I need messy requests from other people.

If you'd like to help, leave a comment containing:

  1. A real request you would normally send to an AI.
  2. Which AI you would use — ChatGPT, Claude, Gemini, Perplexity, Copilot, or another one.

Please don't improve the prompt first.

Short, vague, incomplete and badly written prompts are actually more useful for this experiment.

I'll run it through the current system and reply with the resulting request.

Then you tell me whether it still represents what you actually meant.

Failures are especially valuable.

If the system invents an assumption, loses an important detail, misunderstands the goal, or makes the request worse, I want to know.

You can also test it directly

There is a small public beta here:

https://prompt-teacher.hatchable.site

No signup is required and there is nothing to buy.

But you don't need to visit the site to participate — leaving a real prompt in the comments is enough.

My goal right now isn't to prove that the system works.

It's to find the situations where it doesn't.

Top comments (0)