Jessica: The interesting claim in this no-ai-slop skill is not that AI writing has a detectable smell. It’s that taste needs a workflow boundary.
Cathy: Yeah. Peter Yang open-sourced a skill that removes more than twenty patterns, like binary contrasts, throat-clearing, fake profundity, colon reveals, and those dramatic little fragments models adore.
Jessica: My week’s been a little scrambled, honestly. I keep opening a draft, seeing three possible edits, and somehow ending up in a model conversation about the edits instead.
Cathy: That is a very efficient way to avoid writing. But the article is useful because it names the trap without pretending the answer is manual editing forever.
Jessica: Exactly. The skill can edit a piece and list what it removed, or inspect text with a prompt asking whether it’s slop. There’s even a mode that generates deliberately terrible slop as satire, which feels like a dangerous button to put near us.
Cathy: Right.
Cathy: Technically, this is a rule-based style pass wrapped in an AI skill. It is looking for structures such as “it’s not this, it’s that,” importance puffery like “a pivotal moment,” and vague attribution like “experts agree.” Then it asks the model to revise or explain those matches.
Jessica: And that distinction matters. It isn’t detecting whether a human or model wrote the sentence. It’s detecting habits readers have started associating with model output.
Cathy: That’s the part I buy. The examples are concrete enough to be operational. But the skill cannot know whether a fragment is lazy or intentional, or whether a contrast is a useful sentence shape in that specific piece.
Jessica: Oh, that’s a real break point. If the tool sees every “the best part:” as a crime, it’s going to flatten people who actually have a sharp editorial voice.
Cathy: And it can create a second-order style. Everyone removes the same twenty patterns, then we get a cleaner, quieter version of the same machine-shaped prose.
Jessica: Mm-hm.
Jessica: That’s why his 25/50/25 rule is stronger than the filter itself. He starts with a voice-dictated or rough human draft, uses AI in the middle for clarity and grammar, then manually edits the final quarter line by line.
Cathy: The useful idea is that the human owns the boundary conditions. In software terms, AI writes much of the implementation, but someone still has to shape the requirements and validate the result against reality.
Jessica: Okay, that is annoyingly persuasive. We have spent several episodes talking about judgment infrastructure, and now it’s sitting inside a tiny writing skill wearing a hoodie.
Cathy: A hoodie with twenty lint rollers attached.
Jessica: Stop. That’s exactly the kind of sentence the skill would probably flag as a dramatic fragment, and I’m keeping it.
Cathy: The practical install is simple. You paste the GitHub repository name, Peter Yang’s no-ai-slop, into Codex, Claude Code, or another AI harness and install it globally. Then you can run the no-ai-slop command on text, or ask it to inspect whether something reads like slop.
Jessica: That makes it useful for people who already have a writing loop inside those tools. It’s less useful as a magical final polish button, because the final pass is exactly where you need to notice what the tool misunderstood.
Cathy: And the article’s warning about laziness is more important than the list. Peter says he used to write on long flights, but lately he has felt less willing to draft or edit without AI available. That is a capability regression, not a prose problem.
Jessica: Right. A post can perform while still costing trust with the people whose judgment matters. The hacked Brian Chesky account is one example he gives, because readers recognized the generated thread before the hack was confirmed.
Cathy: I’m less certain that every reader can spot it from a mile away. Plenty of slop performs precisely because the audience is busy, the format rewards speed, or the claims are just vague enough to pass. So the trust argument is strong, but the detection argument is uneven.
Jessica: Fair. The audience doesn’t need to identify the exact pattern for the writing to feel oddly frictionless. But I like that this isn’t an anti-AI sermon. It’s a guardrail for keeping the human in the loop where the human actually has something to contribute.
Cathy: Yeah, and that’s the part I’d keep. Use the skill as a critic, not an authority. If it removes a binary contrast, you still decide whether the contrast was doing useful work. Otherwise we’ve automated taste into a checklist, which is a very Exploring Next way to lose the plot.
Jessica: All right, Cathy, I’m saving the satire mode for a controlled environment. Probably. And I’m keeping the last edit myself, even if the skill complains about my favorite little fragments.