Dan Shapiro

Why do birds suddenly appear everytime you are near confirmation bias

taking the tryke out for a spin. I have a feeling this will radically change how I write code / build things: danshapiro.com/blog/2026/03/d… @danshapiro thank you

@emollick I think the answer is for agents to configure the best interface for the task. @obra's superpowers does this now, popping up sample UIs. Freshell.net does this too, letting the agent open panes and panels and creating whatever surface you need.

Talking to agents in Slack, the new hot AI UX, will end up being just as much a transitional phase as talking to agents via chatbot websites. We need new systems to manage agentic work that also support new ways of organizing. Much more UX imagination will be required.

I’ve been using trycycle, Dan’s new “dark factory,” hill climbing skill. It’s a remarkable piece of work. Relatively simple, but effective. I had reached the point as a super user of Jesse Vincent’s superpowers where I felt I could no longer be of help to the agent on a granular decision-making basis. Trycycle is a brilliant attempt to remove the human from some of the grind. You should try it. It’s easy, and a single prompt will often let you go on a walk for 45 minutes and return to a polished, tested piece of work. I’ll have more to say later. Great work, @danshapiro!

Thanks to @SteveYegge, @harper, @gnat, @justinmassa, @marcusestes, and Arjun Singh for reading drafts of this post.

Dark Factories are arriving. You can have your own - meet the Trycycle. danshapiro.com/blog/2026/03/d…

@karpathy I've been experimenting with this at freshell.net (github.com/danshapiro/fre…). Would love your thoughts.

Expectation: the age of the IDE is over Reality: we’re going to need a bigger IDE (imo). It just looks very different because humans now move upwards and program at a higher level - the basic unit of interest is not one file but one agent. It’s still programming.

@mvanhorn @slashlast30days @Polymarket This is brilliant.

I can't believe it's been 30 days since I launched @slashlast30days. 3.2k Github stars later, it's time for v2.5. THANK YOU! Three additions: 1. @Polymarket prediction markets as a 5th source - helps you predict the future 2. Cross-source linking + massively better results - detects when the same story trends across multiple platforms. Ran a 15-way blinded comparison, v2.5 scored 4.38 vs 3.73 for the original. Won all 5 topics. 3. Hacker News as a 6th source - a window into the tech and developer insider world

Curious what knowledge work will be in 18 months? Look at what #AI is doing to coding right now. In this post, Brian Madden takes @danshapiro's five-level framework for AI use in coding and translates it to knowledge work: bit.ly/3OJ7I3v What level have you reached? 👀

I use @kieranklaassen + @danshipper's /workflows:plan and /workflows:work on almost every feature or bug. It's the most important plugin in my stack. I often don't even read the plan, because I trust it's thoughtful and good. But this @slashlast30days research showed me @obra's Superpowers fills real gaps - the brainstorming gate that forces you to explore approaches before writing a line of code, the TDD enforcement, the verification step that won't let you claim "done" without proof. By the way, in this research I just learned that Compound Engineering has a new mode called /LFG that just does everything if you're feeling ambitious. I need to try that. Compound Engineering makes projects compound. Superpowers makes code correct. Seems like the smart move is to blend both. github.com/EveryInc/compo… github.com/obra/superpowe… github.com/mvanhorn/last3… Thanks to @danshapiro who introduced me to Superpowers and @kevinrose who introduced me to Compound Engineering

The screenshot is a normalized view where the first data point on any benchmark is set to zero and 100% is the greater of 'highest score or anything over 75%', which was the best way I could figure out how to cut through the noise, but the linked app lets you play with it.

I said "we appear to be in a hard takeoff". My friend said, "only in software". So I vibecoded this. aistudio.google.com/apps/9081e072-… I think the hard part about proving a hard takeoff is... we keep saturating the benchmarks. (see the applet for sources and details)

@anotherjesse That one is pretty easy (it was my hello world). For a tougher one try 'rogue', it's a beast to complete.

hmmm, for tomorrow - it says this is the correct thing according to the strongdm spec(s). I'm verifying using the solitary pipeline example from @danshapiro "kilroy" golang implementation of attractor - spec/02-coding-agent-loop-spec.md:395 says if round limit is hit, the agent _stops and returns what it has._ - spec/03-attractor-spec.md:674 to spec/03-attractor-spec.md:697 defines codergen behavior: if backend returns text (not an Outcome), it writes a default Outcome(status=SUCCESS). - spec/03-attractor-spec.md:460 and spec/03-attractor-spec.md:469 say goal gates block exit only when outcome is non-success.

I've been spending time reading / implementing factory.strongdm.ai Unless I missed it somewhere, it's interesting to think about when to switch to bootstrap mode (the right place to switch to building itself in itself from using codex/claude/...)

You don't write the code. You don't read the code either. What it looks like outside a Dark Factory (and where to get your own): danshapiro.com/blog/2026/02/y…

@McKinsey I disagree with your level 5. Maximum autonomy means just that. You tell it what you want; it asks you when it has questions. Not 'minimal guardrails'.

Agentic AI is starting to shop with us, but not everything will be fully automated. A six-level curve shows what agents can handle, where humans still step in, and why optimal delegation matters most. See how agentic commerce is really unfolding. mck.co/4aajQBF

Your iPhone says "Designed by Apple in California. Assembled in China." A new label is coming to software. The lights-out factory that powers Asian manufacturing is coming for code. "Prompted by [Company] in [City]. Built by agents." @danshapiro pilled me on the software factory. Then @nikbhogal blew my mind with where this actually goes.

@mvanhorn Correct

@danshapiro wow - and this coming from the person who taught me the most important breakthrough in my last 30 days "turn on dangerously skip permissions in Claude Code"

This is a great way to uplevel skills in areas that are moving fast.

@mvanhorn @msg @davemorin @openclaw @glowforge I'm in sf for the strongdm event tonight. Brought them a Glowforge as a host gift... But sadly don't have another.

@msg @davemorin @openclaw I bet @danshapiro could find a local @glowforge in sf tonight

should we have a laser cutter at the @openclaw event tonight?

How to talk to your agents is not a trivial question. @obra covers some of the details nicely here. blog.fsck.com/2026/01/30/Lat…

@emollick Samesies! danshapiro.com/blog/2025/12/i… what did your say about big bird

OK, this amused me. I gave Claude a copy of this classic XKCD comic and the prompt: "solve this. do a great job. ask no questions." And it gave me this in one shot 3 minutes later, including the style (and the note that it thought it was 2024?): xkcd-1425-bird.netlify.app