PromptLens
Grammarly, for AI prompts

Type it messy.
Send it engineered.

PromptLens lives in the text box on ChatGPT, Claude, Gemini, Perplexity and Copilot. It catches what will confuse the model before you press send, asks the one question that matters, and ships a prompt the model gets right first time. Fewer re-asks, fewer tokens, better work. On your device. Nothing stored.

10 free prompts to try · mac, windows & chrome · ⌘ ⇧ L
chatgpt.comwatching
Amber · vague request“make it better” gives the model nothing to aim at. Say who reads it and how it should sound.
0checking
Better prompt for Claude · built on your device
<goal>Write a leave request email to my manager</goal> <context>3 days next week, family wedding; Priya covers my work</context> <constraints>warm and professional · under 120 words</constraints> <output_format>Subject line, then the email</output_format>
65% first-shot
of tasks solved by the model's first answer, vs 15% for the prompt as typed. Blind test, 60 conversations.
−59% follow-ups
fewer "no, I meant…" turns per task. That is where the tokens, the energy and the hour go.
0 prompts
run through the Arena, our ML-backed nightly research engine, as of 10 Nov 2026.
₹1 cr/yr
of rework time recovered per 1,000 employees, against a ₹24 lakh licence. See the arithmetic below.
The problem

Every vague prompt is paid for twice.

Once in tokens, when the model answers the question it guessed at. Once in time, when a person spends an hour fixing the answer. Both bills land on whoever is paying for the seat.

4.5h
a week the average enterprise AI user spends revising or redoing AI output. 58% revise regularly; 28% have had AI-assisted work rejected.
Zapier / Centiment, 1,100 enterprise users
41%
of the time, models silently invent the missing requirement instead of asking. The answer looks finished. It isn't.
arXiv 2505.13360
39%
drop in task performance when instructions arrive spread over several turns instead of one consolidated prompt, across 15 models.
Laban et al., arXiv 2505.06120
13%
weekly use of Copilot seats that cost ₹2,495 a month. The seat is paid for; the first attempt doesn't work, so people stop trying.
Larridin

The cheapest place to fix an answer is before the prompt is sent. That is the only place PromptLens works.

Corporate

Save ₹1 crore a year, per 1,000 employees. Here is the arithmetic.

Built from published figures, not ours. Overwrite any input with your own headcount and salary band. The number you change yourself is the one you will believe.

Your numbers

Rework today
108,000 h
a year, across 500 regular users
Hours recovered
16,200 h
≈ the output of 8 extra full-time people
Value of time recovered, per year
₹97 lakh
range ₹52 lakh – ₹1.6 cr at 8–25% rework removed
PromptLens licence
₹24 lakh
per year
Net · return
₹73 lakh · 4×
before tokens, incidents avoided, or lifting the seats you already pay for
AI tokens saved
2.0B tokens
a year: 1.3B from fewer re-asks and capped answers, 720M from slimmed files. Worth ₹82,000–₹8 lakh on usage-billed APIs and agentic tools (Claude Code, Codex, Cursor); nothing on flat-rate seats, where the saving shows up as time instead.

Model, not measurement: the 15% is tied to the Arena's base case (30% of prompts left alone; on the rest, better 42% of the time, worse 22%). The model never falls below the licence cost unless rework reduction drops under 4%. Sources: Zapier/Centiment, BCG, Gallup, BetterUp (~$186 per employee per month lost to low-effort AI output).

  1. Prompt-quality score per teamWhich departments use AI well, which gaps recur, and the trend. The answer boards now ask for: is AI delivering value, not just is it safe.
  2. Secrets caught before they leaveAPI keys, passwords, card numbers, Aadhaar, PAN and client names are flagged and redacted on the device, before the text reaches any provider. Prompt-level DLP, which shadow-AI buyers rank first.
  3. Seats that finally earn their keepPromptLens at ₹400 is the seat that makes the ₹2,495 Copilot seat work: the first attempt lands, so people keep using what you already pay for.
Run a 60-day pilot one department · turns per task and regenerations per task, before and after · the pilot report is your ROI case
Education

Free AI seats are everywhere. Instruction isn't.

AICTE and OpenAI put 150,000 free ChatGPT Go licences on Indian campuses and pledged 500,000 more; Google gave students a free year of Gemini AI Pro. Half of Indian students and educators say they don't know how to get the most out of AI, and 85% feel overwhelmed by it. PromptLens is the training layer that makes those seats worth something, inside the chat box, while the student is working.

01 · learn by doing

Get good at prompting, in the moment

Underlines as they type, with a one-line reason. A 0–100 score. Learn mode explains every fix and tracks a personal skill score across the semester. No lecture, no Coursera tab: the feedback arrives where the mistake is.

Evidence. A 2026 randomised trial in a 979-student CS1 course found every prompting-instruction condition improved skill, and "write, then get feedback on the missing components" gained most and held the gain longest. Prompting skill predicted final-exam score. That condition is Learn mode. arXiv 2602.16033
02 · better work

Fewer fabricated sources, fewer guessed answers

The research template forces source constraints, verification steps and "say if you don't know". Clarifying questions run before the model guesses. The assumptions list makes the AI's guesses visible, which is exactly the habit an AI-disclosure honour code needs.

Evidence. ChatGPT fabricated 38% of references across 400 tested; only 26.5% were fully correct. A 3-hour verification training cut fabricated references in AI-assisted literature reviews. arXiv 2505.18059 · CELTI
03 · industry-ready

A skill score recruiters can see

Students graduate into teams where AI is already in every workflow. A verifiable prompting-skill score and certificate, issued through the placement cell, says they can work that way on day one, instead of spending their first six months prompting badly.

Evidence. 40% of Indian employers prefer demonstrable AI skills or certifications over degrees; 63% increased AI hiring year on year; genAI roles carry a 30–40% base-pay premium. Nasscom–Indeed · EY Future of Pay

A credit-bearing course, ready made

A 30-hour Value Added Course mapped to the 29 rules, with a quiz per module. The tool is the coursework; the score is the assessment. No faculty time.

NAAC / IQAC evidence

A cohort dashboard: prompt-score improvement, most common mistakes, the improvement curve. Usable in accreditation documents.

Placement-cell proof

Per-student skill score and certificate aligned to AICTE's Year of AI and the Nasscom call for credit-bearing genAI courses.

Works in their own language

Write the ask in Hindi or Marathi, get a precise English prompt. Indian-language prompts cost several times the tokens, so this is a real saving on free-tier quotas too.

Honest limit: the trial found gains in prompting skill and its correlation with marks, not a direct lift in exam grades. We sell the skill and the employability signal, not better grades.

Energy, water, carbon

The greenest token is the one you never send.

A follow-up isn't a small message. The model re-reads the whole conversation and writes a second full answer. In a blind test on under-specified tasks, PromptLens cut follow-ups by 59%, tokens by 35% and data-centre energy by 31%. Answering its questions happens on your device and costs the data centre nothing. Slimmed files add to that: the tokens a file no longer carries are never read at all, on the first turn or any later one.

  • 1,200 tokens
    One avoided follow-up. Net of PromptLens's longer first prompt: the re-read conversation plus a second answer, gone.
  • 0.5 Wh
    Energy saved per follow-up (central; 0.17–0.97 Wh across published low/high factors from Google, Epoch AI and Jegham et al.).
  • ≈3 mL
    Water saved per follow-up. About 0.6 mL of cooling water at Google's fleet 1.15 L/kWh, plus roughly 2 mL at the power station that generated the electricity.
  • 0.14 g CO₂e
    Carbon per follow-up at 287 g/kWh (the AWS factor used for Claude). On India's coal-heavy grid, about 0.35 g.
  • 6,000 tokens
    One slimmed file. Half of a ~12,000-token attachment (a 5-page PDF measured 11,751 → 3,685). About 0.6 Wh at 0.1 Wh per 1,000 tokens read (range 0.02–0.25 Wh).
  • 0.8 kWh
    Per heavy user per year (10 under-specified tasks a day and 5 files a week, 250 days): 0.63 kWh from avoided follow-ups plus 0.15 kWh from slimmed files. About 2.5 litres of water and 0.22 kg CO₂e. A little over a kilometre of driving.
7.8 MWh
electricity not drawn, per year
25,000 L
water not consumed (cooling + generation)
2.2 t CO₂e
emissions avoided at 287 g/kWh

Equivalent to 12,000 km of driving, or 16,000 1.5-litre bottles of water.

ScaleEnergy / yrWater / yrCO₂e / yr
One heavy user0.78 kWh≈2.5 L0.22 kg
10,000 users7.8 MWh≈25,000 L2.2 t
1 million users780 MWh≈2.5 million L224 t
Ceiling, follow-ups only: a 3–7% cut across ChatGPT's ~2.5 billion daily prompts24–59 MWh a day28–68 million L2.5–6.2 kt

Modelled, not metered: token counts are measured; Wh per token comes from Google's Gemini serving study (0.24 Wh median prompt), Epoch AI and Jegham et al.; water from Google's published 1.15 L/kWh fleet WUE plus a ~2 L/kWh generation allowance; carbon at AWS's 287 g/kWh. Life-cycle studies (Mistral: 1.14 g CO₂e for a 400-token answer, all-in) suggest the true saving per avoided call is several times higher. Slimmed files are modelled the same way: 5 files a week at ~12,000 tokens, half removed, 0.1 Wh per 1,000 tokens the model no longer reads (central; 0.02–0.25 across the same sources, and lower than the follow-up figure because nothing is written in reply). Each file is counted once, although a chat re-reads it on every turn. Full method, ranges and sources in the First-Shot report. The last row is a ceiling for the whole industry, not a forecast for us.

How it works

Four moves. Under a second.

01 / YOU PAUSE

You type the way you talk, then stop for a second.

PromptLens waits for the end of a sentence or a pause. It never checks on every keystroke and never calls a model while you write.

write an email to my boss about taking leave next week for a few days, make it better and dont make it long but cover everything
31
02 / IT READS YOU

It works out what you meant, from what you already said.

A small classifier picks one of 13 task types. Recipient, tone, length and platform are picked up so you're never asked twice.

write · emailto · your managerabout · leavewhen · next weektone · guessedlength · missing
03 / IT ASKS

One question at a time. Never a guess.

If something essential is missing you get a question with answers to tap. Skip any of them. Nothing is invented on your behalf.

How many days, and what for?
3 days · family wedding2 days · personala weekSkip
04 / ONE CLICK

The better prompt lands in the box. You press send.

Formatted for the model you're on. Undo still works.

<goal>Write a leave request email to my manager</goal>
<context>3 days next week, family wedding; Priya covers my work</context>
<constraints>warm, professional · under 120 words</constraints>
<output_format>Subject line, then the email</output_format>
92
Put in chat box
scroll1 / 4
Before and after

What you typed. What the model needed.

keep scrolling to reveal →
what you typed23

hey can u look at this and make it better, its for some clients, dont make it too long or boring, maybe add a few points about pricing

what PromptLens sends94
<goal>Edit the pasted one-pager for clients</goal>
<audience>procurement leads at mid-size firms</audience>
<constraints>
  under 250 words · crisp, no jargon
  keep every fact and number as written
  add 3 pricing points: starter, team, enterprise
</constraints>
<output_format>Revised text, then a 3-line change log</output_format>
The real thing

This is exactly what you'll see.

01 / you typePromptLens sits inside chatgpt.com. Type as you normally would; underlines appear where the model will stumble.
🔒 chatgpt.com
PL
What can I help with?
31
Needs a little moreAdd these for a better answer:
×
  • Say how long it should be
  • Say what tone you want
  • Pick one: short, or cover everything
Improve prompt
313 tips
Message ChatGPT
↑
31
Needs a little moreAnswer one quick question for a much better answer.
×
What I picked up
EmailFor: your bossAbout: leave next weekTone: professional
How many days, and what for?1 of 2
3 days · family wedding2 days · personalA week
SkipIf you skip: I'll leave the dates for you to fill in
Better prompt
for Claude ▾
Answer the question above and your better prompt appears here. PromptLens asks instead of guessing what you meant.
<goal>Write a leave request email to my manager</goal> <context>3 days next week for a family wedding; work is covered</context> <constraints>Warm, professional. Under 120 words.</constraints> <output_format>Subject line, then the email</output_format>
Details
Put in chat boxCopyImprove with AI👍👎
Put in chat box
What it does

Small fixes inline. One big rewrite for the paragraph.

live underlines

Flags what will go wrong, as you type

Red for contradictions and references to nothing. Amber for vague words, no audience, no format. Hover for the reason and a one-click fix.

structure

Turns a paragraph into a prompt

Goal, context, constraints, output format and success criteria, filled from your own sentences. Formatted for Claude, GPT or Gemini.

model router

Picks the right model, and says why

Quick question, Haiku. Forty-page document, Opus. Always with a cheaper fallback and a rough cost, before you spend a credit.

prompt score

One number you can watch improve

A 0–100 score with a five-way breakdown. The weights learn, every night, from which fixes produced better answers.

privacy guard

Catches secrets before they leave

API keys, passwords, card numbers, Aadhaar, PAN, emails and phone numbers are spotted and redacted in one click.

desktop

Works in the apps, not just the browser

The Mac and Windows apps read the focused text box, so the same card appears in the Claude and ChatGPT desktop apps. Select text anywhere, press ⌘ ⇧ L.

0
rules, all run on your device
0
task types it recognises
0
task-specific templates
0
prompts stored
File slimming

Most of what a file costs is packaging.

An attachment is usually the biggest thing in a chat, and the AI app charges for how it reads the file, not for what the file says. PromptLens turns every file you attach into its cheapest faithful form, one Markdown file for the text and one image file for the pictures, on your device, with no AI call, before it reaches the chat. The same content, a fraction of the tokens.

Why files are expensive

  • A PDF is billed twice. Claude reads the text and a picture of every page: up to about 1,600 tokens a page for the picture alone.
  • Slides are read as pictures. Eight slides of bullet points cost over 10,000 tokens as images and about 600 as text.
  • Web pages and JSON repeat themselves. Menus, scripts and the same field names on every record are paid for on every read.
  • And it is paid again on every turn. The model re-reads the whole conversation, file included, each time you reply.

Nothing is lost quietly

  • Every number is checked. Each number, name, email and link in the original must be in the slim copy. If one is missing, the card turns amber and "Use original" becomes the default.
  • Honest about no gain. Word files and small images save nothing, and PromptLens says so.
  • It never leaves your device. Plain code, no model, no upload. Secrets in a file are flagged before you attach it.
q3-report.pdf5 pages · 332 of 332 numbers kept
11,751 → 3,685 tokens−69%
Slides (PPTX)10,449 → 617−94%
Text PDF11,751 → 3,685−69%
Web page12,301 → 3,797−69%
JSON, 200 records13,674 → 4,353−68%
Spreadsheet (XLSX)6,624 → 4,353−34%
Photo, 12 MP1,600 → 1,049 at 1,024 px−34%
Word (DOCX)already read as textno gain

Also measured: log files 95–100%, Jupyter notebooks 85–93%.

Measured on 7 Nov 2026 with the same content saved in every format, counted with OpenAI's o200k tokenizer as a stand-in. Treat every count as an estimate (±10%); Claude 4.7 and later run about 30% higher. Slimming works before a file is sent: a file already in a chat has to be attached again.

Improve with AI

Want it sharper still? One click, your own model.

The rules engine writes your better prompt on your device, for free, every time. When a prompt deserves more, press Improve with AI and a model of your choice polishes the same structured prompt: richer wording, sharper success criteria, still nothing invented.

  • Off by default. Nothing leaves your device until you click.
  • Your key, free options included. Gemini and Groq have free tiers; Ollama runs fully on your machine.
  • Kept honest. Every AI rewrite is checked against the original structure, and rejected if it changes a single number.
  • Trained by the Arena. The model is shown PromptLens's best-judged rewrites for that task type first.
Gemini · freeGroq · freeOpenRouter :freeOllama · localChrome built-inAnthropicOpenAI
Better promptfor Claude ▾
Rules engine · on your device
<goal>Write a leave request email to my manager</goal> <context>3 days next week for a family wedding; work is covered</context> <constraints>Warm, professional. Under 120 words.</constraints> <output_format>Subject line, then the email</output_format>
<role>You write clear, courteous workplace emails.</role> <goal>Write a leave request email to my manager</goal> <context>3 days next week for a family wedding; work is covered</context> <constraints>Warm, professional. Under 120 words. Lead with the dates. Name who covers my work. End with one clear ask for approval.</constraints> <output_format>Subject line, then the email</output_format> <success_criteria>My manager can approve it in one reply.</success_criteria>
Built on your device. No tokens used.
Put in chat boxCopyImprove with AI
Model router

Stop paying Opus prices for Haiku questions.

The routing table is data, not code. It's re-checked every month against the current lineup, so the advice never goes stale and no vendor is favoured.

  • Four tiers. Quick, light, heavy, novel. Each with a plain-English reason.
  • Always a fallback. The cheaper model that would still do the job.
  • Honest about cost. The structured prompt is longer. The saving is on the answer, which is where the money goes.
The research engine

ML-backed. Agent-driven. Proven again every night.

Every other prompt tool assumes its rewrite helps. PromptLens is the only one that measures it against real model answers, blind, every night, and reverts anything that scores worse. Two systems do that work: the Arena, an ML-backed evaluation and learning loop, and the Lab, a fleet of agentic research bots that test new hypotheses before they are allowed near the engine.

0
prompts run through the Arena as of 10 Nov 2026: real first messages from public chat logs, our own prompts, and hand-written cases across 13 domains.
0
blind judge verdicts. Every judged prompt is scored at least twice with the order swapped; verdicts that flip with position are discarded.
0
consecutive nights of automated evaluation since 24 Oct 2026, on free models and, this week, on a Claude Sonnet judge.
0
rules whose score weights are learned from outcomes, scaled ×0.6–1.4 once a rule has 20+ judged samples.
the arena · ml-backed evaluation loop

Does the rewrite get a better answer? Measured, not assumed.

A nightly pipeline on a pinned corpus: PromptLens rewrites each prompt (a simulated user answers its questions from a hidden intent it never shows the engine), one model answers both versions, a second model judges blind, twice, and a critic labels what helped and what broke.

  • Three-arm measurement. Rewrite vs original, plus a third arm of the original with the same answers in plain words. That separates the value of asking from the value of formatting.
  • Position-swap double judging with paired sign tests; disagreement is a tie, never a win.
  • Self-training classifier. A logistic-regression task classifier (13 task types, trained on ~3,600 labelled prompts) retrains nightly on the critic's relabels, and is promoted only if it beats the old one on a held-out set.
  • Rule-impact learning turns the 0–100 score into "how much better your answer gets if you fix this", and a benefit predictor learns when to stay quiet because the prompt is already fine.
  • Outcome-mined few-shots: rewrites that won outright, kept intent and invented nothing become the examples the AI rewriter sees first.
Latest result. After the Lab's do-no-harm finding shipped on 1 Nov, the loss rate on judged prompts fell from 78% to 24% in a single night (rewrite won 61%, p≈4×10⁻⁷), and 101 of 287 prompts were left untouched, because the best rewrite of a good prompt is no rewrite.
the lab · agentic research bots

Forever improving, one controlled experiment at a time.

Ideas for the engine don't get coded; they get tested. The Lab keeps a ranked queue of hypotheses drawn from the Arena's loss analysis and the prompt-engineering literature, and runs each one as a randomised, blind A/B before a line of it reaches the product.

  • Hypothesis queue with predictions written down in advance ("loss share ≤ 35%"), so a result can fail.
  • Controlled runs: 32 sampled prompts per hypothesis, a sampler check that the arms actually differ, parallel model subagents as answerer, simulated user and judge, and a paired sign test at p < 0.05.
  • Promotion pipeline: a passing hypothesis becomes an Arena candidate, runs as an A/B against the live engine for a trial night and a confirmation night, then is promoted. One change per night; the judge prompt is frozen for the week.
  • Honest bookkeeping: inconclusive results are logged and demoted, invalidated runs are refunded, and every night ends with a regression gate.
Example. H001 tested a lean, do-no-harm rewrite against the per-task-type template: 9–2 in favour of lean. That became the 1 Nov engine. H024 ("restate the length rule last") came back 13–11, inconclusive; it was demoted, not shipped.
The gate that protects everything. After any learning step or promotion, the full suite runs: 71 unit tests, 80 regression prompts (each also checks the do-no-harm invariant) and a 232-prompt context evaluation. If any number gets worse, the night's changes are reverted automatically and the reason is logged. The engine cannot learn its way into breaking something that already works.

How the engine, ML and Arena work →

Why nothing else does this

Templates guess. Rewriters guess harder. PromptLens asks, then proves it.

The "Grammarly for prompts" slot has been empty since the one-shot rewriters faded and the template libraries became paste-boards. Every alternative shares one weakness: nobody checks whether the result got a better answer.

Prompt libraries
& templates
One-shot
rewriter sites
"Improve prompt"
buttons in the model
PromptLens
Works where you already typeChatGPT, Claude, Gemini, Perplexity, Copilot, desktop apps✕✕one vendor✓
Asks instead of guessingone question at a time, tap to answer, skip anything✕✕✕✓
Do-no-harm guaranteeyour words kept; a good prompt is left alone✕✕✕✓
Measured against real answers, nightlyblind, position-swapped, with p-values✕✕internal, unpublished✓
Learns from outcomesclassifier, rule weights and few-shots retrain on what won✕✕?✓
Runs on your device, nothing storedno sign-in, no telemetry, no server sees your promptvaries✕✕✓
Privacy guardkeys, cards, Aadhaar, PAN, client names redacted before send✕✕✕✓
Vendor-neutral model routerright model, cheaper fallback, cost before you spend✕✕✕✓
Token and energy meterthe saving shown as fewer re-asks and shorter answers✕✕✕✓

The moat isn't the rewrite. It's the prompt → rewrite → outcome dataset the Arena builds every night, which no public corpus has, and which gets more valuable as models improve.

Pricing

Try 10 prompts free. Then pick the plan that fits.

Create a free account on this site, download the app (it arrives already logged in) and your first 10 prompts are on us, no card needed. After that, choose Individual, Education or Corporate, each in Pro (unlimited better prompts, 100 file slims a month) or Max (adds Improve with AI and unlimited file slimming). Your score and suggestions keep working on any plan.

individual
₹399 Pro · ₹799 Max / month · $8 / $15 US
  • Unlimited better prompts for Claude, ChatGPT and Gemini
  • Max: PromptLens AI, up to 1,000 AI improves a month, no key needed
  • Chrome extension plus Mac and Windows apps
  • File slimming for every file type, with OCR and bundling (Pro: 100 files a month; Max: unlimited)
  • Privacy guard with one-click redact
Download, then choose Individual in the app
corporate · per seat
₹400 Pro · ₹800 Max / user / month · $10 / $18 US
  • Everything in Individual
  • Privacy guard policies: what must never leave, enforced before send
  • Admin dashboard: usage, quality trend, money saved by routing
  • Team library and house rules
  • Slimming report: tokens saved across the team
  • 60-day pilot in one department, with a before/after report
Download, then choose Corporate in the app
education
₹300 Pro · ₹600 Max / student / year
  • Everything in Individual, with sign-in by college email
  • Prompt Coach: learn mode and a per-student skill score
  • 30-hour credit-bearing Value Added Course, ready for the Academic Council
  • Cohort dashboard for NAAC / IQAC
  • Certificate for the placement cell
Download, then choose Education in the app

Anchors: Copilot ₹2,495 a seat a month; Grammarly for Education about $13–31 a student a year; ChatGPT Edu cleared at $2.50 a user a month. Payment is not switched on yet: during the pilot, plans are given with an access code. Team and campus dashboards are delivered as a report during pilots.

Privacy

Prompts are the most sensitive text you write. It's built around that.

Everything runs on your device

Underlines, score, questions and the rewrite come from a rules engine with no network calls. Your account only tells the app which plan you are on.

Nothing is stored

Prompts are never saved. An optional local log keeps only the task type, score and which rules fired, and never leaves your browser.

AI only when you click

“Improve with AI” sends your prompt to the model when you click it, and at no other time. It is used for that one rewrite and never stored on our servers.

Download

Mac, Windows and Chrome. Five minutes to set up.

Get the desktop app for the Claude and ChatGPT apps, and the extension for the websites. Log in or create your free account here first: the downloads are for account holders, and the app connects to your account by itself. Choose your plan, or enter an access code, inside the app.