System of React Native + Tauri apps: weeks to build (for me). Swift + Gemini Pro 2.5: 12 hours & $80 of cloud credits, yesterday. I’ve never used Swift before.
The new Gemini 2.5 Pro model is insane, but the one prior to gemini-2.5-pro-preview-05-06 was also great. There’s a catch though: make it bring you along, even if you’re kicking and screaming the whole way 馃槄
I’ve adopted a practice of keeping major agent tasks documented in /gen/gen#.md files, first with the planning step (asking it to make mermaid diagrams, which I always seem to have to fix lol please learn how parens need “s someday!) and then adding the implementation deviations at the end.
I’ve even added /gen/gen#.meta.md files so an agent doesn’t need a continually growing context window for major refactor efforts and we can track file by file, or component by component changes in a markdown list.
Just started added gen/future#.md files for future considerations once the app is in a different stage. This mode has felt pretty good, sort of like a step towards GenAI’s version of ADRs.
MCP integration may be next - seems like a standard that these AI-first companies will be adopting for internal tooling. Would be nice to not have to ask for the updates and have it baked into the agent work. Post planning and post task documentation could also follow some format too.
Today I’m sitting on on a pretty major, human and manual refactor so I can better learn the stack.
Would love to know if anyone has any similar practices?