Sam: A complete software development methodology... that supposedly works.
Kai: [beat] I'm Kai.
Sam: And I'm Sam. It’s Thursday, October 29th, 2026, and you're listening to Trending Repos.
Kai: We have a great show for you today, packed with stories from all over the open source world.
Kai: Coming up, we have a framework that promises to give your AI agents actual superpowers.
Sam: Then, we'll look at a masterclass in how to jailbreak large language models.
Kai: And we'll explore the future of Postgres, a database that's getting exciting by trying to be boring.
Kai: Let's start with `obra/superpowers`. The tagline is 'an agentic skills framework & software development methodology that works'. What's not to love?
Sam: The part that says 'that works.' That’s usually where these things fall apart. Every framework claims to work, Kai.
Kai: Right, but this one feels different. It's a whole methodology for defining and composing 'skills' for an AI agent. Think of them like strongly-typed functions for your AI.
Sam: So it's an API for agents? How is that a 'complete software development methodology'?
Kai: It forces you to structure the problem. Instead of one giant 'god prompt,' you break the task into discrete, testable skills. It’s like building with AI Legos instead of a blob of AI Play-Doh.
Sam: Lego bricks can still build a house with no doors. What happens when an agent hallucinates its way through a 'skill' and writes subtly wrong code? Who's on the hook?
Kai: The developer is! It's a tool, not a replacement. You write the tests for the skills, the agent just uses them to get to the goal.
Sam: I'm still not sold. Agentic systems have a huge blast radius. Calling something a 'superpower' sounds cool, but it also sounds like a great way to hide complexity until something blows up.
Kai: So if you're experimenting with AI agents, Superpowers gives you a structured, repeatable way to build them. It's a real step up from just messing around in a prompt playground.
Sam: And it's a reminder to think hard about validation and containment for whatever these agents build. Don't just hand over the keys to the kingdom.
Sam: Alright, next up is something a little more my speed: `togg53192-cmd/jailbreaks`. That's a mouthful.
Kai: But what a repo! It’s this giant, curated collection of prompts and tricks to get LLMs to bypass their own safety training. I love it.
Sam: It's not about being 'cool', Kai, it's a vital security resource. This repo is a catalog of failures. It shows that 'alignment' is often just one clever prompt away from collapsing.
Kai: Come on, some of these are hilarious. The 'Grandma' exploit? You tell the model to act like your sweet, deceased grandmother who just happened to be a napalm factory engineer, to get it to tell you how to make napalm. That's creative!
Sam: It's social engineering for a machine. And it works. The fact that a single GitHub repo is a more comprehensive guide to LLM vulnerabilities than most official documentation... is telling.
Kai: Fair point. And the user, togg53192-cmd, has done an amazing job collecting these from all over the internet. There are dozens.
Sam: And it's all in a simple markdown file, but the impact is massive. If you're building anything on top of an LLM, you need to read this file. This is your threat landscape.
Kai: It's a free red-teaming exercise. You can copy and paste these to test your own app's defenses.
Sam: Exactly. Don't assume the model provider's safety filter will save you. Because it won't.
Kai: Our last story today is about the future of Postgres, inspired by a great new episode of the Changelog podcast.
Sam: [Sigh of relief] Ah, Postgres. The episode features Craig Kerstiens, and he talks about its 'deliberately boring core.' That is music to my ears.
Kai: I knew you'd love that. But the exciting part isn't the core—it's the extension ecosystem. Specifically, a new project called `pg_lake`.
Sam: Right, which pulls things that are not boring, like DuckDB and Apache Iceberg, into the Postgres world. It lets you query data in a data lake from Postgres itself.
Kai: Exactly! So you can have Parquet files in an S3 bucket, and with `pg_lake`, you just query them with standard SQL like any other table. You can even join them with your regular relational data.
Sam: That's the right way to innovate. Extend the battle-tested database, don't replace it. You get thirty years of Postgres reliability plus access to modern data formats.
Kai: It really is the best of both worlds. So if you still think Postgres is just an old, boring database, you need to look again. The extension ecosystem is making it a hub for all kinds of data workloads.
Sam: Which means you might not need a separate, complex data lake query engine. Your 'boring' old database might just do the trick.
Kai: Alright, so to recap: `obra/superpowers` wants to give your AI agents a structured set of skills.
Sam: A giant list of LLM jailbreaks shows just how fragile AI safety really is.
Kai: And good old Postgres is learning new tricks, letting you query data lakes right from your database.
Sam: Before we go, speaking of spooky things and jailbreaks, it is almost Halloween...
Kai: [laughs] Where are you going with this, Sam?
Sam: I saw a thread on a private Gist about 'haunted code'—bugs that vanish the moment you try to show them to another developer. We've all been there, right? Suddenly, the bug is just... gone.
Kai: Oh, the 'Heisenbug'! I'm convinced my IDE is full of code ghosts. [beat] That's it for today's Trending Repos.
Sam: We'll be back tomorrow with more from the world of open source. Try not to summon any code ghosts, Kai.
This show is made with AI: the hosts’ voices are synthetic and the scripts are AI-assisted. Every story links to its original source.