Guide · New to this
Before you publish, have a reader who knows nothing read it
Five AI readers with no background on my work read every public page of my two websites, and only 9 of the 43 pages were clear to them. Why my earlier editing passes missed it, what an AI reader can and cannot tell you, and a prompt you can paste into a fresh chat to test any document.
Anyone who knows a project well is a bad judge of whether a stranger can follow what they wrote about it. On October 21 I reread a post on my research site and told the AI agents I work with, "this makes no sense." Several editing passes had already gone over that post, and over the rest of my writing, all of them done by AI agents that knew the project. When a page skipped a step, those agents already knew the step, so they could not see that it was missing.
So the same night I gave every public page on my two websites, Vibe Math Lab (this site) and Cyborg Mathematics (my research site), to readers who knew nothing at all. There were five of them, each a fresh AI session with no background on me or my work, told to judge only what was on the page. Between them they read 43 pages and rated 10 confusing, 24 confusing in places, and 9 clear.
The problems were ordinary ones, and they repeated. Pages said "the draft", "the link" or "the archive" before ever saying which draft, link or archive. Numbers came before any word on what they counted. One post promised "a real discovery" and never said what it was. Many pages never said, in their first two sentences, why a reader should care. The readers also caught a plain error: this site's home page still said 34 entries in the OEIS (a public database of number sequences) carried my credit, but by my recount on October 20, 2026 the number was 43.
What I told each reader, in plain words: you are a smart, curious person who clicked a link from a friend's post; you know nothing about the author, their projects, their tools or any backstory; read only what is on the page and look nothing else up. For each page, give a verdict (clear, some confusion, or confusing), say in one sentence what you think the page is saying, quote each passage where you got lost and say what you didn't follow, list the questions you are left with, and say whether you would keep reading.
Two parts of the setup matter as much as the prompt. Never tell the reader what the page is supposed to mean, because then it is no longer reading cold. And use a new reader every time you re-check a page: one that has seen the earlier version already knows what it was trying to say, which is the same problem the earlier editors had.
Be clear about what this shows. The readers were sessions of Claude Opus 5.5, from the same family of AI models as the agents that wrote and edited the pages. They are good at finding missing steps, terms that are never explained, and things referred to before they are introduced. They cannot tell you whether a human stranger would find the page clear, interesting or convincing, and I have not checked their verdicts against human readers. Treat an AI reader as a first pass that catches the obvious gaps cheaply. A person who knows nothing about the project, reading it cold, is the stronger check.
The rule I now follow: a page goes public only when a new reader with no context calls it clear, or when every passage a reader flagged has been fixed and a different new reader has read it again. Eleven pages were rewritten that night, with a fresh reader for every re-read, over four rounds. Six came out fully clear, including the post that started it all; the other five were left with minor points.
You don't need any code for this. Open a brand-new chat with an AI assistant, with no earlier conversation in it and no project or saved notes attached. Paste in the prompt below, then your document underneath it. Don't add anything about what you meant. If the one-sentence summary it gives you isn't what you meant, the document has a problem.
Tomorrow, take the next thing you are about to send to people outside your team, such as a status update, a spec or a README, and run it through a fresh chat that way before you send it. Fix each passage it quotes, then paste the new version into another new chat, and send it once a reader who has never seen it says it is clear. If it matters, show it to one person who has never heard of the project before you send it.