Every other writing tool rewrites your sentence. This one refuses.
Eight lenses. One at a time. It shows you what's weak, names the rule, and then makes you fix it yourself.
Paste something you've written and hit a tab
Ask any AI to improve your paragraph and it will hand you a better paragraph. That sounds like the point. It isn't. You get a cleaner sentence and you learn nothing, because the thinking happened somewhere you couldn't see it — and the sentence that comes back sounds like everyone else's sentence, because it was trained on everyone else. Do that for a year and your writing doesn't get better. It gets more average.
What it does
- It never changes a character of your text. Switching lenses changes the highlighting and nothing else. Your words live in a plain text box; the highlights are painted on a layer behind it, so the analysis code physically cannot touch them.
- Eight lenses, worked biggest-first — structure before rhythm, rhythm before words. Each owns a colour, carried as a dot on its tab and again in the highlights it draws, so what you're looking at is never in doubt.
- It quotes the problem and names the rule. "6 flags, mostly jargon. Start with the hedges — they're the cheapest cut." Then it stops talking and waits for you.
- Cutting is the score. Save a version and that becomes your baseline. Points come from words cut × quality held — so gutting the piece earns you nothing, and deleting everything is not a 100% cut.
- It can be told to shut up. ⌥click any highlight to mark that sentence deliberate, and every lens skips it from then on. Your best line is usually a rule broken on purpose.
- 62 detector tests, no API key, no network call. Every score is deterministic. Nothing you type is sent anywhere, because there is nowhere for it to be sent.
An editing tool finishes the job. A teaching tool doesn't.
That's the whole line, and everything in the build sits on one side of it. A tool that rewrites for you is doing your reps. It's a spellchecker for thinking — useful once, useless as a habit, and quietly corrosive if you lean on it, because the judgement never transfers to you.
So this one is built so that it can't rewrite. Not "chooses not to" — can't. The text lives in a plain textarea and the highlighting is painted on a separate layer underneath it. There's no code path from the analysis to your words. That was an architecture decision made on day one, and it's the reason the tool stays honest when it would be easier not to.
What you get instead is a machine that points. It quotes the offending phrase, tells you which rule it breaks and who says so, and hands the instruction back to you. Fixing it is the exercise. That's the bit that makes you better, and it's the exact bit every other tool takes away.
One layer at a time, biggest first
Fixing a word while the structure is broken is wasted work, so the tabs run left to right in the order you should work in.
- Structure
- Pulls out the first sentence of every block and lists them alone. If that list isn't your argument, move blocks before you touch a single word.
- Pace
- One bar per sentence. A flat column means every sentence is the same length. Wide bars are the ones to split.
- Concrete
- Flags the abstract filler — better, seamless, quality, solution — and counts your numbers, names and examples.
- Grammar
- King's version, not Grammarly's: passive voice, hidden subjects, and verbs buried inside nouns. Make a decision → decide.
- Words
- Jargon, hedges, clichés, vague asks, and long words wearing a short word's job. It names the swap. It doesn't make it.
- Adverbs
- Every -ly, with a stoplist so family and reply don't count. Dandelions: one looks fine on the lawn, then there are fifty.
- Voice
- The reads-like-AI scan — delve, a testament to, not just X, it's Y, em-dash pileups — plus a read-it-aloud pass where you click every line you stumbled on.
- Rhetoric
- The only lens that adds. Click one line — your headline, your last sentence — and get fifteen shapes to try on it. Never run it over a whole piece.
The same paragraph, three different questions
Here's a lump of announcement copy of the sort that goes out every day. Nothing changes in the text as you move across the tabs — only what the tool is asking about it.
What it won't do, on purpose
Worth writing down, because each of these looks like a missing feature and is actually a decision.
It won't score your meaning. Semantic scoring past about a thousand tokens returns noise — the same paragraph comes back 2.75 one run and 3.45 the next. So the two lenses that need judgement, Concrete and Voice, hand you the question instead of faking a number. Can I picture it? Is it true or false? Could a competitor sign it? A machine can't answer those. It shouldn't pretend to.
It doesn't know your voice. It catches what doesn't sound human. It can't yet catch what sounds like you — that needs a corpus of your own writing and a model, and it's the obvious next version.
It teaches one school. King, Ogilvy, Harry Dry and Forsyth are all the same tradition: plain-style, Anglo-American, pro-compression. Every one of them argues for cutting and concreteness. None argues for density, ornament or digression — and plenty of good writing is all three. This tool teaches one school well. That isn't the whole of writing, and it would be dishonest to sell it as if it were.
Anyone who has to write and doesn't want it written for them
Founders and consultants writing their own posts, emails and landing pages. People who ship a newsletter and can feel it going flat. Anyone who has pasted a paragraph into an AI, got something smoother back, and had the slightly hollow feeling that they'd outsourced the wrong part.
It's free and it always will be — there's no account to make and no server to pay for. Open it, paste something you wrote this week, and hit a tab.