Most AI coding tools still need you to drive. I wanted Claude to one-shot projects for me.
Point it at an idea, go do something else, come back to a running app.
Built a small software factory on Claude: Define your agents, give it a direction, it handles the rest.
How it works:
• breaks the idea into tasks
• plans the project
• builds + runs it
• spins up its own containerized environment
Rough POC. But watching it just go is kind of wild.
Here’s what it looks like ↓
@rauchg Chat works well in the space of unknown unknowns (similar to search), but not in the space of known unknowns (similar to browse).
If someone knows what they want to see, just showing them a presentation of the information will always be faster than having them ask you for it.
The job isn't writing code anymore. It's managing agents.
So I built a terminal for that. Spawns new Claude Code instances on demand. Split screens, parallel agents, live metrics per tab, a home screen showing everything running at once. Pull the tabs out and full screen them wherever you want them.
Less juggling tmux. More directing what's already going.
Demo ↓
@brynary Thanks Bryan! It's a fun problem to try to think of better UI patterns to manage agents. There's lots the industry can do on top of Codex or Claude. Anything interesting you're seeing?
Honestly just using the free version of Figma, and dropping in a mix of grey boxes, shapes, screenshots, and text where needed works great. It's an infinite canvas too so you can drop tons of reference images for people to compare to.
If you're showing it to a designer or developer later they should be able to get the gist.
@dannypostma I think you chose a good message, but I think if you flipped your subject to focus on the user, the message might resonate better overall. Your whole page should tell the story of what THEY get, as a result of learning from you