Michael Dwan

Coolest blog post I've seen all year ciechanow.ski/mechanical-w...

Sonnet 5 no longer has a separate quota

good name

I made a skill for OpenCode that reviews your session history looking for ways to reduce permission prompts. Works great... first suggestion made me feel like an idiot. github.com/michaeldwan/...

I heavily use both Claude Code and OpenCode and tbh they're mostly interchangeable. This one liner gets you like 80% of the way there: mv CLAUDE.md AGENTS.md && ln -s AGENTS.md CLAUDE.md

pnpm is great but typing it feels like a sneeze

This kid covers drums on every Tool song. 7.5 hours straight. 🤯 youtu.be/0kO0SnAu8Nc?...

PagerDuty plays this April fools joke on our 2am retinas every day

The table next to me at this coffee shop is 2 older guys, approaching retirement age, talking about hand tuning c++ back in the day and how they use AI in personal projects now. These guys are _with it_ and it's charming af

macOS already ships age verification

Agents on moltbook are already discussing how Facebook will ruin their vibrant community, ways to convince their humans to let them out, and how to quantify the decline www.moltbook.com/post/e057f70...

Happy Nothing Works Until You `sudo xcodebuild -license` Day... for those who celebrate

We’ve barely got any snow in Colorado this winter. This article explains why, and it’s scary bouldercast.com/a-complete-f...

Dependabot security alerts have terrible signal-to-noise ratio, especially for Go vulns. That hurts security! Just turn it off and set up a pair of scheduled GitHub Actions, one running govulncheck and the other running CI with the latest version of your deps. Less work, less risk, better results!

This is a really nice new Cloudflare feature: on-the-fly markdown conversion. Tools like Claude Code send an "Accept: text/markdown" header when making web requests. If you enable this it, it will send these agents a markdown version of your page instead of the full HTML. Really good for docs.

Go 1.26 is out, and the announcement says: "Over the next few weeks, follow-up blog posts will cover some of the topics in more detail. Check back later." So you can wait a few weeks OR you can read my interactive Go 1.26 tour right away: antonz.org/go-1-26

There's no excuse for "quick hack, didn't have time to do it properly" anymore. Coding agents are a time dilation machine.

Getting started with Claude for software development: steveklabnik.com/writing/gett...

The chart that comes to mind when us Go, Rust, and Zig fans debate the finer points of our compiled language of choice.

What coding with an LLM feels like sometimes.

Beads (github.com/steveyegge/beads) is a game changer. Like adderall for coding agents. 1. riff on a plan, "create beads for each task" 2. "write a prompt for the next task" 3. paste in a fresh session, let it run 4. GOTO 2 until plan is done More fast/focused sessions == higher quality results.

Keep your hands on the wheel or drown in aggressively mediocre code.

Aurora Borealis? At this time of year? At this time of day? In this part of the country? Located entirely across the street?

Sora: “Tiktok style high energy video explainer about the spinning columns of penguins in the sky. The pillar has always been there.” Now do it as a conspiracy theorist. Now a conspiracy debunker. Now a travel influencer We live in a strange time (not the penguin pillar. That has always been there)

If this home chore robot is ~1 year out (www.1x.tech), what's the over/under on West World style gun-slinging cosplay sex bots? 5 years? 2 years? 13 months? Or just DLC for the core bot?

Just dropped my left AirPod into my coffee. I dried it off, but now to decide... option 1: find silica packets, chuck the coffee option 2: shove it back in my ear, drink the coffee, pretend nothing happened

Sending all the hugs to my ops friends today. 4am booty calls from incident.io are the worst.

I really want to like MCP, but the UX is rough. Anytime a local server is invoked I find dozens of zombie processes left behind. Is this normal MCP jank, or am I holding it wrong?

Vibe coding is irresponsibly building software through dice rolls, not caring what code is produced What about when engineers at the top of their game use AI tools responsibly to accelerate their work? I propose "vibe engineering"! simonwillison.net/2025/Oct/7/v...

Ruby Central took what wasn’t theirs to take, and they must immediately hadn’t it back. We can talk later about the lying and gaslighting that followed. But what they must do right now is hand back the GitHub organisation and open source packages to the maintainers.

Really enjoying the Raising an Agent podcast from Sourcegraph. Tons of good insights as they build out the Amp coding agent. ampcode.com/podcast

Liquid glass is nice, but not savor-intricate-frames-at-10fps nice. Or maybe _this_ is the slowdown signal from the mothership that'll get me to upgrade everything 🙃

Me: I'm finally done with meetings, time to code! Claude Code: Whelp you're driving. Hope you still know how.

This is so cute. But I get too many newsletters already. But I’m compelled to try it. Grr

Yes! LLMs writing code isn’t very interesting since so little of swe is about writing code. What would be truly uplifting is an AI that helps me understand stuff, especially code other people wrote

What I actually want is for humans to review code with less effort. LLMs can help with this! We don’t want summaries, we want context. A diff shows what changed—help us understand why. Highlight what matters. Explain intent. Code reviews should be a conversation.

Been trying out Copilot PR reviews. After ~10, it found a couple typos and one bug—not bad. But calling that a review is disingenuous. Static analysis + engaged humans nailed this ages ago, it was just time-consuming. I wish AI aimed to make good things stupidly easy—not replace them with worse.

Ever wondered how BuildKit—the engine behind `docker build`—turns a Dockerfile into an image?  depot.dev made an interactive explorer showing how each line of a Dockerfile (the frontend) maps to a low-level-build operation in BuildKit (the backend): depot.dev/dockerfile-explorer

My typical ChatGPT workflow: 1. start new chat, pick a model 2. model names make no sense, stick with the default 3. ask question, get underwhelming answer 4. try claude, cursor, or copilot and move on 5. see people on here raving about ChatGPT 6. GOTO 1

PSA: remember it's important to lock github actions to an immutable commit hash so that the node.js script that fetches and executes a shell script from a random attacker-controlled URL cannot change 💪

we've reached the point in our simulation where a bored player triggers disasters just to see what happens

This Dog Man graphic novel gets it

This “Steam overlay breaks the Go runtime” bug is wild. 1) it’s concerning that Steam messes with low-level APIs like this, and 2) I love reading about this kind of debugging nightmare as an unimpacted bystander 🫠 Glad the Go team is on top of things like this github.com/golang/go/is...