A SKILL.md only helps if your coding agent can find it. I built skill-land to simplify that setup. Run `npx skill-land` to install a skill the agent can read. I’m building small tools for AI agent workflows and sharing what I learn. Follow along if you’re working on the same problems. How do you manage skills across coding agents?

Before asking your AI coding agent to fix a bug, tell it how you’ll know the bug is fixed. For example, with a login bug: Provide the failing request. Point it to the relevant authentication code. Define the expected response. Have the agent reproduce the failure first, fix the underlying issue, and rerun the same check. A fix isn’t verified just because the agent says it’s done. Inspired by Anthropic’s guide on context engineering for AI agents: anthropic.com/engineering/ef…

Been thinking about a problem in AI-assisted development that deserves more attention. A recent paper got me thinking about how we build with coding agents, so I wrote my first X Article.

I agree with Tim. Game development is more accessible than ever. Anyone with an LLM and an idea can now start building a game. We’ve had access to the same game engines for years, yet the quality of games varies wildly. AI will accelerate development, but it won’t magically give everyone good game design instincts, creativity, or taste. And calling a game “AI slop” simply because AI was used to make it doesn’t make sense to me. If a game is fun, I’ll play it regardless of how it was made. Not every game needs to appeal to everyone. What matters is whether it delivers a great experience to its intended audience. AI will be part of game development. There’s no doubt about that. At the end of the day, players remember the experience, not the tools used to create it.

Came across a paper with a funny name today 😂 FDE-Gym. But the problem it tackles is pretty serious. AI agents can write code, but can they figure out what the customer actually needs when the requirements are incomplete? The researchers tested 13 model configurations on 95 simulated delivery tasks. Even the best configuration resolved only 14 out of 95 under the benchmark’s scoring setup. My takeaway: Working code ≠ successful delivery. The real challenge? Understanding what nobody thought to specify. Paper ⬇️ arxiv.org/abs/2610.22350

OMG NO WAY 😱 I LEGIT THOUGHT THIS WAS AI AT FIRST?? LIKE WHATTT 🤯

Today’s pick : AI memory has a trust problem: a saved summary can leave out the detail that matters later. A recent paper explores building memories from retrieved evidence, keeping track of their sources, and returning to the original conversation when the memory isn’t enough. Memory should help AI remember and show its work. arxiv.org/abs/2609.34438

A new model for free every week… OpenCode is making me remember the oldest rule on the internet: If you’re not paying for the product, you might be the product. 😂 So… what happens to our data?

Blanked reset and please no global resets

is Bumble even still a thing ??

Today’s pick : ChunkAct: Action Chunking for Agentic Robotics (arXiv preprint, submitted Oct. 13). Robots may not need an LLM call for every move. ChunkAct groups routine actions together and calls the model when a decision or failure needs attention. On LIBERO-Pro, the authors report 5 points higher task success than the previous state of the art, with 2.9× fewer LLM calls. arxiv.org/abs/2610.18874

Lmao 🤣

Could a botnet of $55 Android TV boxes run an AI model? Anthropic tested a 26B model split across 15 boxes. It ran slowly, at 0.39 tokens/sec. Separately, a 9B model got inference running on 8 of 9 hardware types tested. They didn’t demonstrate a real-world AI worm. The security question is what happens when vulnerable devices are connected. anthropic.com/research/self-…