Under the hood

An engine, a few small models, and a loop that keeps score.

PromptLens is built in three layers. Each one has a different job, and each one checks the one before it.

The pipeline

From a pause to a better prompt, in about 40 milliseconds.

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.

You pauseend of a sentence Check29 rules · underlines Understandtask · your clues Decideask, or rewrite Structurefrom your own words Formatfor Claude, GPT, Gemini Score · model · costshown in the panel Privacy guardsecrets caught first Improve with AIoptional · your own key ← the only optional network call
01 · The engine

It reads your prompt the way a careful editor would.

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.

red · confuses the modelmake this shorter but cover everything

References to nothing, contradictions, instructions that only say what not to do.

amber · weakens the answerwrite something good for a few people

Vague words, no audience, no length, no shape for the answer.

privacy · caught before sendingmy key is sk-live-9f3…

Keys, passwords, card numbers, Aadhaar, PAN, emails and phone numbers.

check

29 rules, each with a reason

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.

decide

Asks only when it must

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.

structure

About 70 kinds of task

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.

what you typed

write an email to my boss about taking leave next week for a few days, make it better

→
what the engine understood
taskemail
foryour manager
aboutleave, next week
toneprofessional (guessed from "boss")
lengthshort, by default
askhow many days, and why?
02 · The models

Small models that tune the engine, never replace it.

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.

what job is this?

Task recognition

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.

which advice matters?

A score that learns

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.

should I speak at all?

Knowing when to stay quiet

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.

Recognises the right taskon prompts it was not trained on
v0.390.9%
v0.497.4%
v0.598.7%
Asks about something you already saidlower is better
v0.317.8%
v0.43.3%
v0.51.4%
Prompts with any needless naglower is better
v0.335.1%
v0.47.8%
v0.53.0%
Your details carried into the rewriterecipient, tone, length, platform
v0.339.8%
v0.490.8%
v0.592.4%
03 · The Arena

Does the rewrite get a better answer? Ask the models.

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.

fair

Judged without knowing which is which

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.

explained

Every loss has a reason

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.

real

Real prompts 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.

night one

It caught a bug no one had noticed

"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.

!

It can't learn its way into breaking things

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.

Try it

See it on your own prompts.

Free, five minutes to set up, and it runs on your device.