# How to avoid AI slop: start with what a human actually said Source: https://aaronmakelky.com/blog/how-to-avoid-ai-slop Last updated: 2026-11-13 Ryan Doser and I talked about AI slop on AI Rabbit Holes. Why volume stopped working, the anti-slop skills and design rules we use, and why the source of your content decides whether it's slop. November 13, 2026 · 11 min read AI Rabbit Holes · AI Slop · AI Writing · Content Repurposing · Claude Skills · DESIGN.md · SEO [All my AI Rabbit Holes episodes](/ai-rabbit-holes) ![Value, not volume. How to avoid AI slop, with Ryan Doser. A grid of identical faded cards that each say Write me a blog post surrounds one white card labeled Worth publishing: starts with what a human actually said, with three steps, a real conversation, your skills and gates, and iterate on the results.](/images/blog/how-to-avoid-ai-slop/hero.png) ![](https://i.ytimg.com/vi/DV2LiXVKzCc/hqdefault.jpg) Ryan Doser and me on AI Rabbit Holes: how to avoid AI slop. Don't blame the AI tool. Blame the human that's using it. That's what I always tell people. Claude can be used for whatever you prompt it to do. Ryan Doser and I spent an episode of [AI Rabbit Holes](https://www.youtube.com/watch?v=DV2LiXVKzCc) on AI slop and how to avoid it, recorded November 6, 2026. Our episodes go up on [Ryan's YouTube channel](https://www.youtube.com/@RyanDoserAI). [Ryan](https://ryandoser.com/) brought his anti-slop skill and his repurposing process, and I brought my writing audit and my DESIGN.md. It follows our episode on [the best free GitHub repos for non-coders](/blog/best-free-github-repos), where both of those first came up. ## Slop means low effort [Watch this part of the episode (4:25)](https://www.youtube.com/watch?v=DV2LiXVKzCc&t=265s) First I have to shout out the pig farmer in Canada who told me he hates that slop has been weaponized as a negative word. For me it's all about being low effort. Use blogging 10 years ago as an example. There was human slop blogging, where people were trying to get backlinks or affiliate links and the game was mass-produced content. But you had to write it yourself. Those people produced human slop like it was a BB gun. It took some manual effort, it wasn't super powerful, and it didn't scale well, but they played the volume game. LinkedIn had it, Twitter had it, blogging had it, TikTok had it before AI could edit videos. The difference now is you have a chain gun that can crank out stuff at basically unlimited scale. If you want to do low effort and win at volume, it's a million times easier, faster, and cheaper than it was three to five years ago. You can't win on volume anymore. I don't know any marketing strategy where doing more of the crappy thing would make it good. There's a minimum level of volume, of course. You need blogs, you need videos, whatever that is. But you can't say if I scaled it times 10, I'd get times 10 the result. Everybody wants to be Gary Vee and do 300 videos a day. That's not a winning formula for 99.9% of people, and it's not worth the tokens you'd burn to get there. Try to win on value, not volume. [![Same volume game, bigger gun. The BB gun, human slop 10 years ago, shows six squares: you had to write it yourself, it was not powerful and did not scale well, and it still played the volume game. The chain gun, AI slop now, shows a block of 140 squares: basically unlimited scale, easier, faster, and cheaper, same low effort. Win on value, not volume.](/images/blog/how-to-avoid-ai-slop/bb-gun.png)](/images/blog/how-to-avoid-ai-slop/bb-gun.png) Same low effort, a lot more of it. Ryan's definition is that AI slop is human slop at scale. Mediocre copy, design, and editing existed long before generative AI, and he thinks the best models are already better than a mediocre copywriter, designer, or editor. The other half of his definition: someone opens ChatGPT or Claude, says write me a blog post or an email about this, and copies and pastes the output with no skills and no context behind it. He also pointed out that Gary Vee and Alex Hormozi aren't pulling 300 pieces a day out of thin air. Their teams repurpose a huge library of their own long-form podcasts and videos. ## Don't start a faceless channel [Watch this part of the episode (10:28)](https://www.youtube.com/watch?v=DV2LiXVKzCc&t=628s) Would you start a faceless AI-for-marketing TikTok or Instagram channel today? Or would you attach your name, your trust, and your own experiences, and teach through that? Anybody else can out-automate a faceless channel, and all the content in it is going to be random slop repurposed from what other people say. It's a losing battle, even if you win. Ryan pulled up the faceless channel courses live. He says the people selling them make their money from courses, communities, affiliates, and sponsors, not from the channels. And remember, those people are getting monetized off teaching you how to do the thing, not off the thing they're teaching. [![What the faceless channel course sells, and what pays. What you're sold: a faceless AI channel, automated posting, and content at volume. Where the seller's money comes from: courses, communities, and affiliates and sponsors.](/images/blog/how-to-avoid-ai-slop/faceless-money.png)](/images/blog/how-to-avoid-ai-slop/faceless-money.png) Ryan's breakdown of where faceless channel courses make their money. ## AI content isn't automatically slop [Watch this part of the episode (12:00)](https://www.youtube.com/watch?v=DV2LiXVKzCc&t=720s) Ryan's take is that people who say AI content sucks just admitted they don't know how to use it. Content made with dialed-in skills, good context, and an original long-form video behind it is a different thing from typing write me a beginner's guide to GitHub into a chat box. You know what people who wrote with pencil and paper said? That's keyboard slop. You're just typing all those words. It's the same idea. That might be the medium you create things through, but that doesn't make it slop. No one today would say, you just made that with the internet. Of course I did, everything I do is connected to the internet. In five years people won't say it about AI either. [![The medium doesn't make it slop. A timeline from pencil and paper, to the keyboard, where people said that's keyboard slop, to the internet, where nobody says you just made that with the internet, to AI, where in five years nobody will say it either.](/images/blog/how-to-avoid-ai-slop/mediums.png)](/images/blog/how-to-avoid-ai-slop/mediums.png) Pencil, keyboard, internet, AI. ## Keep an anti-slop skill, and keep changing it [Watch this part of the episode (13:18)](https://www.youtube.com/watch?v=DV2LiXVKzCc&t=798s) Ryan's anti-slop skill is a markdown file in his VS Code setup, full of things he's told it not to do: the if-not-this-then-that lines, words like delve, robust, and utilize, em dashes, too many emojis. The rule in his AGENTS.md is that every written output, from a blog post to a YouTube title, goes through an anti-slop and brand voice gate. A second model reviews it and gives a ship or no ship verdict. Nothing publishes on the first draft. My tip is to always update yours. What tuned our brains into AI slop a year ago aren't the same phrases, design styles, and color schemes we see today. New models still write in predictable patterns, just not the same ones as the last models. This is why [GitHub](/blog/github-for-marketers) is so good for it. Any time I see one, even in other people's writing, and think, I know this is AI, it grinds my gears, I take a link to it or copy and paste it and tell my agent: put this in my anti-slop skill, I never want that type of output. There are two levers you can pull. You can put up guardrails and say what not to do. Or you can proactively nudge it: I write like this, use these examples, this is my sentence structure and vocabulary. So you might say no delve, but I like to use analogies. Any good skill should have some of both. [![Two columns titled Two levers for a writing skill. Don't do: no delve, no inflated language, and any pattern that makes you think I know this is AI, added the moment you see it. Do: I like analogies, this is my sentence structure and vocabulary, and examples of how I actually write.](/images/blog/how-to-avoid-ai-slop/two-levers.png)](/images/blog/how-to-avoid-ai-slop/two-levers.png) Tell it what to avoid, and show it what you sound like. Mine is a public skill called [AI Writing Audit](https://github.com/a-makelky/ai-writing-audit). I think of it as the negative, what to avoid in the writing. It's not an AI detector. It looks for the patterns our brains associate with low-effort AI writing, like inflated language and certain words. Everybody's top AI word is still delve, and the funny thing is I don't think models use it much anymore. They used to a ton. [![Aaron Makelky's public GitHub repo ai-writing-audit, which checks written text for common signs of AI-generated writing, suggests fixes, and creates a change log](/images/blog/build-or-buy-saas/ai-writing-audit-repo.jpg)](/images/blog/build-or-buy-saas/ai-writing-audit-repo.jpg) My [AI Writing Audit](https://github.com/a-makelky/ai-writing-audit) skill. Hit fork to add your own patterns. It audits the text, tags the patterns it finds, corrects them, and gives you a change log. If you're a coder, that's a diff. If you're a writer, it's a red line, like your editor coming back with notes or your English teacher with a red pen. It shows you what it found and what it would change, and it gets the low-hanging ones out of your writing. If you know what you're doing, hit fork and add your own patterns or change things. Everybody should have some sort of gate on their written content. Even if you took the transcript of Ryan talking, AI is going to inject some AI writing patterns into it, unless you tell it to turn the transcript into a blog word for word, which is not what you want. ## Design slop and DESIGN.md [Watch this part of the episode (16:28)](https://www.youtube.com/watch?v=DV2LiXVKzCc&t=988s) Ryan showed the [slop page from Impeccable](https://impeccable.style/slop/), which catalogs the design habits of different eras and models: purple gradients, glass panels, beige editorial layouts. He keeps it behind the scenes when he vibe codes a landing page so the agent steers away from them. I use Impeccable too. [![Impeccable's Slop page: AI slop changes with the era and the model. A catalog of 68 design tells sorted into typography, color and contrast, layout, motion, copy and imagery sits beside a 2022 example with purple gradients, glass panels and neon glow, flagged for an overused font and low contrast text.](/images/blog/how-to-avoid-ai-slop/impeccable-slop.jpg)](/images/blog/how-to-avoid-ai-slop/impeccable-slop.jpg) [Impeccable's slop page](https://impeccable.style/slop/) sorts design tells by era and by model. If you really want to be pro level, here's what helped me. I have a hook in my agent so any time it does design, it automatically has to call that skill. I don't have to say, okay, draft one is done, now use Impeccable and add some guardrails and suggestions. It just must use it. If you don't know how to set that up, ask your AI tool. For his slides, Ryan starts from his brand guide, with his colors, fonts, and text, then adds [HyperFrames](https://github.com/heygen-com/hyperframes), HeyGen's free GitHub repo, for the visuals and motion graphics. What I tell people is you need a DESIGN.md. It might start with landing pages, but I use mine for [video](/blog/codex-hyperframes-video-editing-workflow), images, and web. If you build apps you can use it for apps too. Tell your agent, I need guidelines so my visuals are consistent. If you're thinking, I'm not a designer, I don't know what typography is, you don't have to know. Have it ask you. Show it examples, give it a site you've made that you want to copy, or a font you like. [![The colors block of Aaron Makelky's DESIGN.md file, with named colors ink #000957, ink-soft #2A3170, ink-muted #555A7A, cobalt #073CCB, white surface, pale blue surface #F4F7FF, two line colors, Electric Teal #2DD4BF, a teal wash, and a destructive red](/images/blog/how-to-avoid-ai-slop/design-md.png)](/images/blog/how-to-avoid-ai-slop/design-md.png) The colors in the DESIGN.md that runs this site. If you went to Claude and said, put a landing page together for me and make it like my website, you wouldn't get the same output. I'm not a designer. I don't understand half of what's in mine. But when I get an output, it's consistent, it looks the same, and it's exactly what I want without having to prompt it. When I add new pages, it also makes sure everything uses one of my components. Part of that is because Ryan taught me about Cloudflare's [Is It Agent Ready](https://isitagentready.com/) scanner, and this site scored 33%. I took the prompts from it and had my components include what I needed to score higher. Then there's a list of do's and don'ts, so any time an agent is going to put something out, it knows what to do and what not to do. Ryan also has a brand foundation builder skill that scrapes a client's website and social profiles and writes four or five markdown files: a brand guide, an ICP guide, an offer guide. Include a writing voice in there. If it's any kind of team or company and not one human's personal brand, everybody needs one, or AI is going to go way off the rails from what you want. ## Start with what a human actually said [Watch this part of the episode (18:29)](https://www.youtube.com/watch?v=DV2LiXVKzCc&t=1109s) Ryan's foundation is long-form video, a podcast like ours or a long tutorial. The transcript holds your original thoughts and words, and that's what goes through the skills and gates. I think the most underused content from a long-form video is the transcript, the words, the markdown file. The most overused is clipping. Here's a pro tip: I use clips from long-form, including this podcast, but between the time and effort to repurpose and reformat them, I often find that if I had a good tip, re-recording it on my own gets a better output than cutting it, putting captions on, and posting it. I've done that for companies and for myself, and it's a volume game. Three years ago you had to be tweeting and TikToking every day. Now I usually get better results recreating something from the long-form video, using the transcript. Look at Ryan's post on the 10 best GitHub repos for marketers. Could I open a new chat right now and say, write me a blog post about the 10 best GitHub repos for marketers? Could Claude do that? It could make one up. It could pick the repos and write the copy. Could I then say, put it on my blog, and email my list? Yes. So why is Ryan getting utility and giving value from his? It starts with the original source of the content. It wasn't Claude. It was the transcript of us, the humans, talking. These are the repos Ryan uses, which are not the same ones a Google search or Claude would pick. There's an element of taste, and the same goes for mine. If the original piece came out of an AI tool, everything you get out of it is going to be slop. You're chopping up slop into different shapes. Next step: take a real transcript, like this one, and say, hey, write a blog for me, with no skills and no specificity. Claude can do that too. But everybody's done that. Ryan's has diagrams, tables, rich media, and subheaders. If somebody took the same YouTube transcript and just said write a blog, that's slop, because you're firing the cannon thinking, I need 10 blog posts. [![Three ways to make the same blog post. One: ask Claude to pick the repos and write it, and the source was AI, so it's slop chopped into different shapes. Two: paste a real transcript and say write a blog, with no skills, and it's still slop. Three: a real transcript plus your skills and gates, then improve it from the results, and it's worth publishing.](/images/blog/how-to-avoid-ai-slop/three-ways.png)](/images/blog/how-to-avoid-ai-slop/three-ways.png) Same topic, three starting points. If you start with a good foundation where the original content is human, you have the opportunity to repurpose it in a valuable way, especially when it's based on original human output like the transcript of this episode. But each additional step, like turning it into a good email newsletter, takes some skill. ## How Ryan does it [Watch this part of the episode (20:28)](https://www.youtube.com/watch?v=DV2LiXVKzCc&t=1228s) From one episode, best free GitHub repos for non-coders, Ryan's one slash command produced [10 best GitHub repos for marketers](https://ryandoser.com/best-github-repos-for-marketing/) on ryandoser.com. He says it took him close to a year to refine. The command: 1. Embeds the video and optimizes the H1, H2s, and H3s, with a table. 2. Has Codex generate an on-brand diagram with his logo. 3. Spins up subagents to find each GitHub repo and take screenshots, then embeds them with captions and alt text. 4. Adds internal and external links. 5. Runs the post through his anti-slop skill and an anti-spam skill built from Google's spam policies. [![Ryan Doser's post 10 Best GitHub Repos for Marketers in 2026 (All Free), opening with a short answer and a Watch section that embeds our Best Free GitHub Repos for Non-Coders episode](/images/blog/how-to-avoid-ai-slop/ryan-github-repos-post.jpg)](/images/blog/how-to-avoid-ai-slop/ryan-github-repos-post.jpg) The post Ryan's slash command built from our GitHub repos episode. He searched it live in an incognito window. It ranked first for best GitHub repos for marketing, and our video showed up in the videos tab. His Blotato review video, which he says he hasn't touched in over a year, is still the top video for that keyword, and ryandoser.com is the top cited source. He does the same for shows he guests on. Every video also becomes an email: the same diagram from the blog post, the video summarized in about five bullet points, a link to the post, and a call to action for whatever his offer is. Ryan says Google doesn't penalize AI content. It penalizes spammy AI content, which its [spam policies](https://developers.google.com/search/docs/essentials/spam-policies) call scaled content abuse: many pages generated to manipulate rankings without adding value for users. Ryan turned those guidelines into the anti-spam skill. He walks his [community](https://www.skool.com/ai-marketing-insiders) members through the process step by step. [![Google's spam policies section on scaled content abuse: many pages generated to manipulate search rankings and not help users, no matter how it's created, with examples starting with using generative AI tools to generate many pages without adding value for users](/images/blog/how-to-avoid-ai-slop/google-scaled-content-abuse.jpg)](/images/blog/how-to-avoid-ai-slop/google-scaled-content-abuse.jpg) Google's line is value, not the tool: [scaled content abuse](https://developers.google.com/search/docs/essentials/spam-policies#scaled-content) applies no matter how the pages are made. ## If you won't iterate on it, don't do it [Watch this part of the episode (34:47)](https://www.youtube.com/watch?v=DV2LiXVKzCc&t=2087s) A thing I think about a lot with slop: if you're not going to iterate on something, don't waste your time doing it. If you're going to run the same blog skill every week on your recordings and never change it, you might as well not do it, because it's going to keep being slop. If you take the results, the traffic, the click-through rate of your emails, whatever it is, and say next week I can do better, I can tweak this and refine it, it won't be slop for long. Ryan closed by saying to call out human slop when you see it, because it never went away. He also said the volume play used to work, and plenty of people haven't adapted since. If you're not going to keep up your blog or do better with your content, and you think, but if I just put stuff out every day or every week I'll win at volume, it's a complete waste of your time. You'd be better off not doing it at all. If you want help setting up your own writing or design gates, [book a 30-minute call](https://cal.com/aaron-makelky/intro), or follow me on X at [@theaaron](https://x.com/theaaron). ## Related pages - [All field notes](/blog) - [The best free GitHub repos for non-coders](/blog/best-free-github-repos) - [Our GitHub for marketers episode](/blog/github-for-marketers) - [Build it or buy it: the SaaS tools I'd replace with AI](/blog/build-or-buy-saas) - [The Lovable tax, and the website stack I use instead](/blog/lovable-alternative) - [My Codex and HyperFrames workflow](/blog/codex-hyperframes-video-editing-workflow) - [Your AI app should not own your work](/blog/your-ai-app-should-not-own-your-work) - [All AI Rabbit Holes episodes](/ai-rabbit-holes) ## Sources and references - [Watch the episode on YouTube](https://www.youtube.com/watch?v=DV2LiXVKzCc) - [Ryan Doser](https://ryandoser.com/) - [Ryan Doser on YouTube](https://www.youtube.com/@RyanDoserAI) - [Ryan's best GitHub repos for marketing](https://ryandoser.com/best-github-repos-for-marketing/) - [AI Writing Audit on GitHub](https://github.com/a-makelky/ai-writing-audit) - [Impeccable's AI design slop page](https://impeccable.style/slop/) - [HyperFrames by HeyGen](https://github.com/heygen-com/hyperframes) - [Cloudflare's Is It Agent Ready scanner](https://isitagentready.com/) - [Google Search spam policies](https://developers.google.com/search/docs/essentials/spam-policies) - [Ryan's AI Marketing Insiders community](https://www.skool.com/ai-marketing-insiders) - [Aaron on X (@theaaron)](https://x.com/theaaron) ## Common questions ### Is AI content the same thing as AI slop? No. Aaron Makelky compares it to people who wrote with a pencil calling typed work keyboard slop. AI is the medium, and the medium doesn't make it slop. Slop is low effort. He tells people not to blame the AI tool but the human using it. ### What should an anti-slop skill include? Two levers, according to Aaron. One says what not to do, like no delve, and gets every pattern that makes you think "I know this is AI" added the moment you see it. The other nudges the model toward how you do write: your sentence structure, your vocabulary, examples of your writing, or the fact that you like analogies. He says any good skill has some of both, and that it has to keep changing because new models write in new predictable patterns. ### Why does the source of the content matter for AI slop? If the original piece came out of an AI tool, Aaron says everything you make from it will be slop: you're chopping up slop into different shapes. If the source is a transcript of real people talking, like an AI Rabbit Holes episode, the picks and opinions in it carry taste a Google search or Claude wouldn't pick. Even then, pasting the transcript and asking for a blog with no skills still gives you slop. ### Does Google penalize AI content? Ryan Doser says no. Google penalizes spammy, mass-produced AI content, which its spam policies call scaled content abuse. Ryan turned those guidelines into an anti-spam skill that his blog posts go through, alongside his anti-slop skill, and his AI-assisted posts rank on Google. ### Does Aaron use a DESIGN.md for anything besides websites? Yes. He uses it for video, images, and web, and says it works for apps too. He isn't a designer and says he doesn't understand half of it, but his outputs come out consistent without extra prompting. His agent also has a hook that calls the Impeccable design skill automatically any time it does design work.