PromptLens is built in three layers. Each one has a different job, and each one checks the one before it.
Finds what will trip the model up and builds the structured prompt. No network, no account.
02 · THE MODELSSmall trained models decide what kind of job this is, which advice matters, and when to stay quiet.
03 · THE ARENAEvery night, real AI answers decide whether the rewrites actually got better results.
Everything on the solid path runs inside your browser or desktop app. The dotted box is the only step that can use the network, and only when you click it.
The engine is fast, predictable and private by design. It runs the same way in the extension and the desktop apps, so an improvement made once reaches every surface.
References to nothing, contradictions, instructions that only say what not to do.
Vague words, no audience, no length, no shape for the answer.
Keys, passwords, card numbers, Aadhaar, PAN, emails and phone numbers.
Every underline comes with a one-line explanation of why it hurts the answer, and either a one-click fix or a quick question. Nothing is flagged without a reason you can read.
Missing the thing the prompt is about? You get one question at a time, with answers to tap. Missing only a detail of form, like length? You get a sensible default you can see and change. It goes quiet after you dismiss it a few times.
An email, a cover letter, a bug report, a logo brief and a translation each need a different shape. PromptLens recognises the job and gives the rewrite the length, format and success test that job needs.
write an email to my boss about taking leave next week for a few days, make it better
Rules decide what to say. Trained models decide which job this really is, how much each piece of advice matters, and whether to say anything at all. They're small enough to run in your browser and they improve from every night of judged results.
Is "average of these numbers" a calculation or an image request? Is a list of features a spec or an edit? Everything downstream depends on getting this right, so a trained model works alongside the rules and learns from its mistakes.
Some fixes reliably lead to better answers; others barely matter. The score's weights shift over time toward the advice that has been shown to help, so a higher score means a better answer, not just a tidier prompt.
Some prompts are already good. A third model predicts when a rewrite won't help, and PromptLens simply says the prompt reads clearly. Tools that nag get ignored.
A rewrite can look tidy and still make the answer worse. So every night PromptLens tests itself on the only thing you care about: the answer you get back.
The judge sees only the original request and two unlabelled answers. It judges twice with the order swapped, and if its verdict changes with the order, the result is thrown out. The judge is never the same AI that wrote the answers.
A separate critic writes down what the rewrite got right, what it broke, and how PromptLens should change. Reasons are counted across nights, so the most common mistakes rise to the top and get fixed first.
Most of each night's test set is prompts real people actually typed, messy and half-finished, in English and Hinglish. Results are always reported separately for real and synthetic prompts.
"average of these: 23, 45, 12, 67" was being treated as a request for a picture. Every calculation was misread the same way, and the rewrite looked so tidy nobody spotted it. The arena's first verdict did. It was fixed the next morning.
After every night of learning, PromptLens runs its full test suite. If anything that worked yesterday works worse today, the whole night's changes are undone automatically. And nothing reaches the version on your computer until a new release is deliberately made.
Free, five minutes to set up, and it runs on your device.