It has a UI to manage projects and the "catalogue" of capabilities the Rouge has to build with. Building a project has two phases: Spec involves some or all of the following, depending on how big the project is: brainstorming, competitive research, taste evaluation, specs (writes up the user stories the loop tackles one at a time), infrastructure decisions (for the foundation step), design system, legal docs, marketing copy. Each of these disciplines has its own prompt, outputs, and quality gates.
Build is then an autonomous build/evaluate/fix loop until it ships. I took inspiration from Karpathy's autoresearch for the build loop. Rouge takes one story at a time, builds it, and then tests it when built. The work is tracked on disk between these loops.
The tests are external browser QA, cross-checking specs are completed so the individual LLM doesn't judge itself, the external checks do. There's a fix loop if something breaks so it fixes itself before progressing to the next story and escalates to the user if it can't fix itself.
At the end, it uses a gstack inspired skill to do a product walk (without the specs) and give blind feedback in the way a first customer would.
What has it built? github.com/gregario/openfleet is a fleet management software with maps, maintenance scheduling, trip management, driver management (and a different driver app) and GDPR features for tracking vehicles only in office hours. It cost $24.
github.com/gregario/construction-coordinator is a self build management tool with supabase auth, push notifications, material dependency tracking, daily briefs, a gantt chart for upcoming work, and snag list tracking. It was 27 user stories and built overnight! It cost $49.25.
Where is the project at? It currently invokes ClaudeCode subprocesses to do the build, it can run via API key or via your ClaudeCode subscription. The dashboard is a bit rough, the chat interface needs work to be polished. It has a big red banner in the dashboard and on the readme saying it uses --dangerously-skip-permissions (I'm working through changing this to --allowed-tools). There's cost management, but its not bulletproof... It could technically blow up your credits or delete your file system.
Something that is not there yet is the Rouge should be able to build new capabilities for the Rouge. If a project spec says the project needs a graphing library and that isn't a capability, it should first spin up a project to build that first. It would be really cool as an open-source project if a user could add a new capability to the Rouge for the community to use when they build something new.