Ryan Singer

Product & process. Creator of Shape Up. Prev: 37signals. [email protected]

[Ryan Singer @rjs]

@joshpitzalis https://t.co/fAT5Gs7PaU

[Ryan Singer @rjs]

A little hard to explain without seeing all the details, but Meso has a Lispy code-as-data step in it, which is what allows it to get better at breadboarding each time I drive it from the agent. It rewrites a certain part of itself to improve its renderer. So there are challenges to figuring out what a deployed version looks like.

[Ryan Singer @rjs]

Working on a side project with Amp. But the first version that came back was too far from what I wanted. So I shifted to a breadboard view and iterated from there. Like moving up to higher altitude. Here's a timelapse of 24 revisions. In the end, pushed the new plan to Amp.

[Ryan Singer @rjs]

This corresponds to what Taguchi called system design vs. parameter design. When the wider system isn't right, I can't whack-a-mole my way to a better one. Need to step back and do whiteboard work at high altitude. Then after the system is closer, I can tune the million dials to get it more exactly where I want it.

[Ryan Singer @rjs]

Did the first draft with CC + Meso (the app on the right). That generated the shape and plan.

Pushed the docs to a repo.

Opened the repo in Amp. Asked it to build. It asked me a few questions, revised the plan, and built it.

Clicked around on the resulting build in Amp. Realized it was too far away from what I wanted in too many ways.

Pulled the repo back to CC + Meso. Did the iterating you see above. Got to a point where the breadboard more concretely described what I wanted. CC revised the plan the whole time as it generated each breadboard.

Pushed the revised plan + shape back to the repo. Asked Amp to pull and one-shot it.

Now I'm clicking around in the Amp portal and the core UI flows are more 80% what was in my head. Now I'm able to see the more "local" problems / changes I want to make all over it.

[Ryan Singer @rjs]

Echoes of Lisp here. With agent-generated code, the code is data. So we can do things like have deterministic paths that are themselves rewritten by feedback loops.

[Ryan Singer @rjs]

@marchattonhere I found that skills and prompts alone aren't consistent enough -- the agent will make up a new or slightly different way to do it each time.

So I built an app the agent uses to express the breadboards in a fixed language, with a fixed set of components, in a deterministic renderer.

Then there's a feedback loop where the agent rewrites the components it's using in that language, to do things we didn't think of before or to follow corrections.

[Ryan Singer @rjs]

@marchattonhere I found that skills and prompts alone aren't consistent enough -- the agent will make up a new or slightly different way to do it each time.

So I built an app the agent uses to express the breadboards in a fixed language, with a fixed set of components, in a deterministic renderer.

Then there's a feedback loop where the agent rewrites the components it's using in that language, to do things we didn't think of before or to follow corrections.

[Ryan Singer @rjs]

@marcinperlak Also good to know — this is how the agent revised the plan. The breadboard was our common view of the plan.

[Ryan Singer @rjs]

@marcinperlak That process of iterating on the breadboard was 100% agent-driven, with Claude Code on my machine. I am prompting on the left side.

[Ryan Singer @rjs]

@kencytbb No this is new tooling.

[Ryan Singer @rjs]

Inner loops and outer loops both have to navigate search space. ie. there is no final factory.

[Ryan Singer @rjs]

For fewer things to do more you have to be clearer about what they're meant to do.

[Ryan Singer @rjs]

Another word for "simplicity" is "leverage."

How much of the same thing can I get out of fewer elements.

[Ryan Singer @rjs]

"Interface" is such a powerful word because you can use it talking about software or you can use it talking about a value chain. And the meaning doesn't change.

[Ryan Singer @rjs]

Another word for "simplicity" is "leverage."

How much of the same thing can I get out of fewer elements.

[Ryan Singer @rjs]

Said differently, value is at the interface.

[Ryan Singer @rjs]

"Product" means working backward from an interface.

The interface -- human, API, CLI -- is where the value happens.

[Ryan Singer @rjs]

"Product" means working backward from an interface.

The interface -- human, API, CLI -- is where the value happens.

[Ryan Singer @rjs]

Fresh

[Cameron Langford @cplangfo]

loving cognition's new team page

[Ryan Singer @rjs]

@neejkapoor I fussed with so many AI-driven diagramming tools, and in the end, Claude Code does a better job inventing diagrams from scratch given clear guidelines for the look and behavior I want. I'm learning to lean into that.

[Ryan Singer @rjs]

I've been talking a bit about CAD for product design. Here's an example of what I'm talking about. This is from a client project. First we worked on a whiteboard to get to a concept for the project. It's a mix of wireframes and pasted screenshots ...

[Ryan Singer @rjs]

@neejkapoor No, it's their comments feature.

I tried to somehow integrate AI with TLDraw a few times in the past. Came to the conclusion that TLDraw's unique strength is actually how extraordinarily good the human interface is. So I rely on it for that -- when I want to draw something myself -- and use other tools for generating diagrams.

[Ryan Singer @rjs]

This was so cool to see. Every time my Stratechery invoice arrives, I cheer a little.

[Ben Thompson @benthompson]

I always intended Stratechery to have a sponsorship component; it just took me 12 years to get back to it. That I could do so with the company that made the subscription business possible makes me so happy. There is nothing better than endorsing a product you whole-heartedly believe in.