People are going AI-blind. Here are the tools that'll still get you noticed.

On 22 August, a post called "I'm becoming AI-blind" (opens in a new tab) reached 357 points and 356 comments on Hacker News (opens in a new tab) in about a day. Rafal Cymerys had written four minutes' worth of an observation about his own reading, and several hundred engineers turned up to say the same thing had been happening to them.

The observation might make you feel slightly uncomfortable, yet familiar. As written, he kept opening documents at work and finding that he could not process them. Not disagree with them. Not skim them. His brain would move over the words without picking anything up, and then he would send the sender a question the document had already answered.

He sat down and worked out what the unreadable documents had in common - it was AI.

Slop in action

He gives three examples:

• A design document that reads like a paste from Claude, carrying Claude-specific analysis and phrasing: "This cuts just through it", "The first gate is real".
• A twenty-page marketing deck that mixes a reasonable strategy with technical architecture that means nothing, and pitches itself with "It's not selling X, it's selling Y" and "The Redis backbone redefines the product".
• A requirements document that describes a simple concept at length, in prose that reads like someone's internal reasoning rather than a decision.

The general flow of sentences and the attempt to pitch every small detail as a breakthrough quickly give it away. If your document describes the checkboxes in an RBAC configuration view for an enterprise application, don't sell it like you've just invented fire.

He says he feels pre-trained on AI-written LinkedIn posts, emails and websites. Others in the thread compared it to banner blindness, which is the right comparison. Nobody decided to stop seeing display ads. The skipping happened below the level of choice, because the sheer volume made it necessary.

The article closes on something that we seem to be finding more and more: the same AI that was supposed to make him more productive is what slows him down now.

AI-blindness, defined

AI-blindness is the learned reflex of skimming past text that carries the surface markers of low-effort AI generation, whether or not the content underneath is any good.

And this is the whole problem: our reflex fires on the markers, not the substance. A useful document written by a person who happened to run it through a model for tidying gets skipped alongside a worthless one.

There is an argument that this cannot be happening, because most research finds humans are poor at identifying AI-generated text. But the argument against this is that nobody is claiming to spot a carefully edited piece: it’s the unedited default output of a general-purpose model, which now has a house style that several hundred million people have read several hundred times each.

Why this costs solo builders more

If you work at a company with a known name, AI-blindness is an annoyance. Your document gets skipped, you have an awkward thread, but the work still gets done.

If you’re one person with a product nobody has heard of, the reader has nothing to override the skip with. No logo they trust, no employment contract that makes them obligated to read it, no prior piece of yours they liked. Every cold email, every launch post, every home page starts from zero and has about a second to prove it was written by someone who actually thought about the problem.

That second is now being spent on pattern-matching for slop rather than on your words.

It applies to the visual surface too; the default output of an AI coding agent has a look: Inter for everything, purple-to-blue gradients, cards inside cards, gray text on colored backgrounds, that left-hand border, a rounded-square icon tile above every heading. Paul Bakaus's Impeccable (opens in a new tab) names those as explicit anti-patterns, and it is the same phenomenon as the one involving words. A visitor who has seen forty of those pages immediately recognizes an AI website as soon as they land.

The AI tells agreement

What might be surprising is how much agreement there is about what the tells are.

Wikipedia's WikiProject AI Cleanup has been reviewing new submissions for undisclosed AI content since 2023. Out of that work came "Signs of AI writing" (opens in a new tab), a field guide of roughly 15,000 words built from patterns the volunteers kept finding across thousands of flagged articles. It is deliberately descriptive rather than prescriptive: observations, not rules.

It catalogues overblown symbolism, promotional tone where the register should be neutral, repetitive transitions, rule-of-three phrasing and editorial commentary a human editor wouldn’t have written.

The open-source writing skills draw on the same evidence base. The anti-ai-slop-writing skill screens for 50+ banned words, 35+ banned phrases, 16 banned sentence openers and 10 structural patterns, and cites Carnegie Mellon research from 2025, the Wikipedia guide and Buffer's analysis of 52 million posts as its sources.

The skills that strip the tells

These are free, readable and small enough to tweak to your liking (or implement into your own Skills, which I’ve done):

anti-ai-slop-writing (opens in a new tab) by Jalaaldeen. MIT, 173 stars. Constraints on vocabulary, structure, punctuation and formatting, plus a category most of the others miss: accuracy failures, meaning invented statistics, fabricated quotes and fake anecdotes. Cross-compatible, and the SKILL.md works pasted into a plain ChatGPT or Claude conversation if you have no harness at all.

stop-slop (opens in a new tab) by Hardik Pandya. Separate phrase, structure and rhythm checks. This is the one I lean on; my own tone-of-voice file cites its references directly, and the rhythm check catches things the word lists cannot, like four consecutive paragraphs of the same length.

unslop (opens in a new tab) by The Clay Method. MIT, 22 stars, and the most interesting of the five. Two modes, audit-only or rewrite, four voice presets, and an eight-criterion rubric scored out of 40 with a pass mark of 32. The Python scripts run standalone, including one that validates that every number, name and date that persists after a rewrite.

Its best feature is wiki_sync.py. The rules are partly derived from Wikipedia's Signs of AI writing page, and the skill can re-check that page, diff it against the last sync and update its own banned-phrase list with whatever the Wikipedia editors added. A detection tool that maintains itself against a living public corpus? With the rate that new models are released, it doesn’t get better than that.

It also exempts quoted examples by default, so a tutorial does not flag its own bad examples, which is the false-positive handling everyone else should replicate.

anti-slop-writing (opens in a new tab) by Adewale. Framed as review, draft and rewrite rather than a filter, which suits editing something you already wrote.

Impeccable (opens in a new tab) by Paul Bakaus. Apache 2.0, 61.1k stars. The visual half. 59 deterministic detector rules, 23 commands, and a CLI that runs the rules with no LLM and no API key at all. It started from Anthropic's frontend-design skill and added the detection layer. 10/10 no notes, my absolute favourite.

comedy-writers-room by Tals Raviv; an honorable mention for this one, which isn’t really related to slop, but it makes things funny. Because if there’s one thing LLMs struggle greatly with, it’s being funny.

Intent capture

Every skill above is a filter; it runs after the text exists and removes the parts that read generated. That works, though it doesn’t save a written doc that’s beginning from scratch.

The more interesting pattern showing up across the skills repos is intent capture: making the agent state what it is trying to do, in writing, before it produces anything, then checking the output against that statement.

Impeccable does this for design. Before any code, the agent writes a five-block contract into the artifact's opening comment, 150 words at most: thesis, own world, story, first viewport, and form. Then the build gets checked against its own stated intent. It also picks direction from an outside seed rather than the model's favorite, so theoretically you move that “three rotating fonts on a cream background” issue.

Vercel's ADR skill (opens in a new tab) does the writing equivalent. It uses Socratic questioning to capture intent before drafting, presents a structured summary of what it captured, and won’t move to the drafting phase until a human confirms or corrects that summary.

Addy Osmani's interview-me (opens in a new tab) sets the test that I think makes the most sense: if you cannot write the desired outcome in one sentence right now, the ask is not clear, and you should not be drafting yet.

What I do

I run Personal Voice (opens in a new tab) first, which is the one skill in Romy's free set (opens in a new tab) I would not write without. It builds a private writing profile out of things I actually wrote, so the model has a point of view to start from rather than a house style to fall back on.

Then the anti-slop pass, then a read-aloud and manual edit, then publish.

The open question I have not answered is whether the reflex generalizes. Right now readers are skipping the surface markers of one model generation, and those markers change every few months.

If AI-blindness turns into a broader distrust of anything polished, the advantage moves to writing that is visibly rough, (what does this mean? Spelling mistakes? Poor grammar?). I don’t know what you do with that, except keep writing as many first drafts as time allows.