The 'anti-slop' files fighting AI slop are AI slop

The sharpest AI post on X last week was not about a model. It was about a markdown file.

Ethan Mollick, the Wharton professor whose feed has become required reading for anyone using AI at work, posted at @emollick that he had been reading one of the viral "anti-slop" instruction files people feed their AI agents to stamp out generic writing. His realization: the file itself was almost entirely AI written, and using it would, in his phrase, "hyperstition your outputs into mega-slop." The fix for slop was made of slop. His prescription was blunt: you need "a sense of taste," or someone who has one, and then you should write your own instructions.

The post stung because of how many people it described. Here is the context, and why I think he is mostly right.

The cottage industry of magic files

Over the past year, a whole configuration-as-markdown economy has grown up around AI agents. Tools like Claude Code, Cursor, and Codex read plain-text files, with names like CLAUDE.md, AGENTS.md, and SKILL.md, that tell the agent how to behave. So people started trading the files themselves like recipes.

The famous example: in January, developer Forrest Chang distilled Andrej Karpathy's observations about where AI coding goes wrong into a single CLAUDE.md file with four behavioral rules. That one markdown file, no code at all, has collected more than 100,000 GitHub stars, which puts it among the most-starred repositories in history. Since then, "anti-slop writing" skill files have gone viral too, promising to strip out the telltale AI habits, the hedging, the stock phrases, the em dash pileups, across a dozen different agent tools.

Mollick's point is not that these files never work. It is that the moment they went viral, people stopped writing them and started generating them. Ask an AI to produce a comprehensive guide for not sounding like AI, paste in the result, and you have outsourced the one part of the job that was yours.

Why the fix makes it worse

"Hyperstition" is a strange word doing precise work. It means a fiction that makes itself true. Applied here: an anti-slop file is usually a long list of banned cliches and bad patterns, so every time your agent reads its instructions, its context fills up with vivid examples of exactly the writing you are trying to avoid. You are rehearsing the failure. And when the file was itself AI generated, every rule in it is phrased in the very voice it claims to ban.

There is a second-order problem Mollick did not spell out but which follows directly. When thousands of businesses paste the same viral file into their agents, their output converges on the same corrected style. You have not escaped generic. You have upgraded to a newer, shared flavor of generic. The Karpathy-derived file works precisely because it is short and encodes real judgment. The failure mode is the 400-line downloaded monster nobody at the company has actually read.

Taste is the scarce input

This is a theme Mollick keeps circling. Earlier this month, in a post at @emollick about Anthropic's Fable 5, he argued the right way to use a frontier model is to "start by asking for everything and find the shape of the frontier from there," rather than the small, cautious requests that suited older models. Different subject, same underlying claim: the model is no longer the limiting factor. The operator's judgment about what good looks like, and the ambition to ask for it, is.

That matches what I see in practice. Two companies can run the identical model with the identical viral config file and get wildly different results, because one of them knows what a good customer email, a good quote, a good follow-up actually reads like, and edits toward it.

What this means for your business

If you run a business and you have been collecting prompt templates and agent files from X or YouTube, Mollick's post is the warning label. The downloaded file is a starting point at best. The instructions that actually work are the ones written from your standards: how your company talks to customers, what your process really is, what you would never send. That is authorship work, not download work, and it is worth an afternoon of any owner's time.

It is also the part we spend the most time on when New Face Design builds automation for Fox Valley businesses. The free process audit exists for exactly this question: not which magic file to paste in, but what your version of good looks like, and how to wire an agent to hit it every time.