stop-slop: A Skill File That Strips AI Tells From Your Prose

Hey everyone, welcome to the Builder's Briefing for May 28th, 2026. I'm Alex, joined as always by Sam.

Hey! Good lineup today. We've got a text file that's taking over GitHub, some real talk about AI product-market fit, and DuckDuckGo is having a moment.

Yeah, let's jump right in because the top story today is honestly kind of delightful. The highest-engagement item on GitHub right now isn't a new framework or a model — it's a skill file called stop-slop. Over thirty-three hundred engagements and climbing.

I love this so much. It's literally a curated set of instructions you drop into your coding agent's context that says: stop saying 'delve into,' stop saying 'it's worth noting,' cut the hollow enthusiasm. Just write like a person.

Right, and the creator Hardik Pandya clearly struck a nerve because everyone has felt this pain. You read something and you just know it's AI-generated even if you can't pinpoint exactly why. This file gives you a way to define what your output should not sound like.

What's wild is the bigger pattern here. I think within six months, AI style guides are going to be as standard as your linting config. You'll have a file in your repo that says 'here's how our AI writes' right next to your ESLint rules.

Exactly. If you're shipping anything with user-facing text — docs, emails, in-app copy — fork this, adapt it to your brand voice, and add it to your system prompt. The competitive moat isn't using AI, it's using AI that doesn't sound like AI. Link in the briefing.

Alright, speaking of AI getting real — Simon Willison had a sharp take this week.

Yeah, Simon argues that Anthropic and OpenAI have crossed the line from impressive demos to genuine product-market fit, specifically with coding and productivity tools. And his warning to builders is pretty direct: if you're building wrapper products, the platform companies are now your direct competitors, not just your API providers.

That's a tough pill to swallow for a lot of startups. And there's a related piece from TechCrunch about tech CEOs suffering from what they're calling 'AI psychosis' — leadership making resource decisions based on hype cycles instead of actual user needs.

The term is provocative but the observation is real. If you're a technical leader pushing back on AI-for-everything mandates, that article gives you vocabulary for the conversation. Also worth flagging — PostHog published an incredibly honest walkthrough of training their own custom models instead of just wrapping APIs. Best cost-benefit breakdown from a real product company I've seen this month.

Oh, and one nerdy gem — there's a resurfacing finding from 2024 that GPU matrix multiplications actually run faster with quote-unquote predictable data. The structure of your input data affects matmul performance in ways you might not expect.

Right, so if you're doing custom inference or training and you're about to throw more hardware at the problem, maybe benchmark your data patterns first. Could save you some money.

Let's talk dev tools. A couple things caught my eye.

So there's a new incremental parser called Handy — designed for coding tools that need to re-parse on every keystroke without choking. If you're building an IDE plugin or an AI code assistant, this could be a really interesting alternative to tree-sitter for specific use cases.

And then there's Rosalind, which is a Rust genomics toolkit that runs whole-genome pipelines on a laptop. No cluster infrastructure needed. The bioinformatics angle is cool, but the real story is Rust making previously server-class workloads viable on local hardware.

That pattern keeps showing up everywhere. Also, someone built a full book publishing pipeline using Git, plain text, and open-source tools — completely bypassing Adobe and Microsoft. If you manage any document-heavy workflow, it's a great template for version-controlled pipelines.

On the infrastructure side — GitHub had another outage, right?

Yep. PRs, issues, Git operations, the API — core workflows went down. If your CI/CD, your deployments, or your agent tooling depend entirely on GitHub's availability, yesterday was your reminder to build graceful degradation. Or at least have a plan.

Also Cloudflare dropped something called Flagship, which looks like it extends what you can deploy on their edge without managing origin servers. Worth checking if you're building on their platform.

Alright, two startup stories worth highlighting. First — DuckDuckGo search traffic is up twenty-eight percent after Google's aggressive AI mode push. Users are literally voting with their clicks.

That's a real number. If you're thinking about SEO strategy or alternative discovery channels, this trend matters. Google pushing AI answers is actively pushing users away.

And then there's a detailed post about Stripe being too friendly to so-called friendly fraud — their dispute process systematically favoring buyers over merchants even when merchants have clear evidence. If you're running SaaS or a marketplace on Stripe, understand those chargeback dynamics before they cost you.

Ooh, that one hits home. Alright, quick hits?

Yeah, rapid fire. Someone compressed all of human cooking into two megabytes — a fascinating data compression paper on recipe representation. There's a developer who went from Rust back to Ruby and wrote about why. And the hiring-without-whiteboards list is trending again on GitHub.

Love that list. Also saw a piece about C array types being weird — pointer and array decay still tripping up experienced devs. Some things never change.

And there's a heated thread on Hacker News — two hundred seventy comments — about big tech's anti-labor playbook coming for Wikipedia. Links for all of these in the briefing.

So here's the takeaway for today. The most popular thing on GitHub right now is a text file that makes AI less annoying. That tells you exactly where the pain is shifting — from 'can AI do this' to 'can AI do this without embarrassing me.'

Right. If you're building with LLMs, invest in output quality as aggressively as you invest in capability. Add style guides, add negative examples, add tone constraints. Treat polish as a product feature, not an afterthought.

The builders who get that right are the ones who win the next round of user trust. That's it for today's Builder's Briefing. We'll be back tomorrow — until then, ship something good.

See you all tomorrow!