Steno
A product, not a repository. Steno is a writing standard that stops AI prose announcing itself — sold as a one-time purchase, with the rules, the copy, the brand, the landing page and the payment flow all built by the person who wrote the rules.
The problem
Everyone can tell when something was written by a language model and almost nobody can say why. It is not grammar. It is inflated openers, the rule of three, participial padding, uniform sentence rhythm, em dashes at twice human density — a dozen habits the model reaches for when it has nothing specific to say.
The advice that exists is vague: “write naturally”, “add personality”. Steno names twelve families of tells and gives each one a concrete substitution, so the fix is mechanical rather than a matter of taste you either have or do not.
How it works
Two artifacts. An instruction layer that sits in custom instructions, project instructions or a system prompt, so the model writes clean text from the start instead of you fixing it afterwards — shipped in four versions for ChatGPT, Claude projects and styles, a short variant for character limits, and a per-message variant for any chat. And a skills file that runs as an editing pass over text you already have, in the Claude Skills format, working as a plain prompt if you do not use skills.
Decisions worth defending
- 01
Two hard constraints hold the whole thing together.
Nothing invented, nothing dropped. An edit changes how the prose reads, never what it says. That single rule is what makes the standard safe to run over someone else's copy, and it is why the editing pass returns text unchanged when the text is already clean instead of finding something to do.
- 02
Tiered rules, not a flat checklist.
Fixing em-dash density in a paragraph that has no stance is polish applied to a broken thing. The tiers force structure and stance first and leave cosmetic tells last, so the pass cannot produce prose that is locally clean and globally empty.
- 03
No claim about beating detectors.
Detector scores move and none of them publish their method, so any promise about them is a promise you cannot keep. Steno targets the part a writer can actually control — the patterns that make a human reader stop and think a bot wrote this. Refusing the easy marketing claim is the reason the honest one is worth something.
- 04
No app, no login, no subscription.
The product is text. Payment runs through Buy Me a Coffee — card or UPI — files unlock on the page as soon as it clears, and a copy goes to email. Wrapping an account system around it would have added a database, a support surface and a churn problem, for no benefit to anyone buying it.
- 05
The site is written in the standard it sells.
Three long-form guide pages double as the SEO surface and as the proof. If the copy read like AI output, the product would be its own counter-example — which is a useful kind of pressure to put on your own landing page.
Proving it works
The demonstration is the landing page, and it is falsifiable in one click.
- A live before/after the visitor can toggle. On the full sample: 11 tells found, 118 words became 51, every claim kept.
- Hand-written HTML and CSS, no build step and no dependencies — original SVG logo mark rendered out to a full favicon and app-icon set.
- Structured data for Product, FAQ, Article and breadcrumbs, plus a sitemap and crawl rules that admit assistant bots.
- The Open Graph card is generated from source HTML rather than hand-edited in an image tool, so it regenerates instead of drifting.
- Deployed on Vercel with clean URLs, an explicit cache policy and security headers.