I've watched a lot of smart, technical people write prompts. They read the guides, learn the frameworks, and do everything right: role, context, constraints, output format, numbered rules. The prompt comes out clean and well organized.
Then the output is just okay. It's stiff, and it misses the point in some way that's hard to name.
My take after about six months of building a prompt tool: the prompt usually isn't wrong. It's too literal. It follows the rules and loses the common sense.
Think about a job interview
The best candidates don't recite their resume. They figure out what the interviewer actually cares about and speak to that. They back up what they say with one real example instead of ten adjectives. And they're easy to work with, so you want to keep talking to them.
A good prompt does the same three things:
- It speaks to intent.It covers what you're actually trying to get done and why, not just the task.
- It proves it with an example. One concrete example of good output does more than a page of rules.
- It's easy to work with.Plain language. It doesn't stack twelve rules on the model for something that needs three.
Over-engineered prompts usually fail the first one. They're so busy with structure that they never say what good looks like in this specific situation.
The example that made it click
Two people I know were building an AI feature that summarizes metrics. Their prompt was solid on paper. But hidden inside it was an assumption: trending up is always good.
That's true for revenue. It's not true for costs, churn, or risk. The model was doing what it was told, and the summaries suffered for it.
No framework catches that, because nothing in the prompt is structurally wrong. You catch it by asking what a coworker would ask: “Is up always good here?”
They ran the prompt through Meerkat. The new version could tell when down was good and up was bad, depending on what was being measured. They tested both on real data, and the new one gave better answers.
Sometimes the right move is to ask
A lot of AI tools are built to one-shot everything: you paste something, they rewrite it, and you're done. When the prompt is missing something important, the rewrite just fills the gap with a guess. Then you refactor, then refactor again, then give up.
The fix is the same as with a person. If something important is unclear, ask. If it's clear, don't ask. Just do the work. And if the prompt is already good, say so and leave it alone. Over-polishing a prompt that works is its own way of making it worse.
So I built Meerkat around that
Meerkat (getmeerkat.dev) is a prompt tool built on this idea. It runs your prompt past Claude, GPT and Gemini, asks a question when something important is missing, and gives you back a prompt that reads like a person wrote it. It's still a work in progress.
The free roast is the fastest way to see what it does. Paste a prompt you use at work, no signup needed.
If you've got a prompt that “should work” but doesn't, I'd like to see what it catches. And if it tells you your prompt is fine, believe it.
Roast a prompt you use at work.
Free, no signup needed.