People tend to optimize their own productivity automating reviews and flooding PRs with code that looks decent. For multithreading it’s a bug wakamole. Maybe engineering authorship will move to iterating on specs, who knows.

Example code: gist.github.com/janodev/92df...

On that note, a very useful prompt is ”what would have helped you to know to solve this problem faster?” You can add the resulting information to the code, a blog post, or project documentation to supplement LLM shortcomings and for your own learning too.

I’ve been involved in project estimation lately. It’s massively shorter using Claude Code, even at current price. It’s beyond me why Apple doesn’t use Deepseek to clone CC and give it away as an onsite free alternative to boost hardware sales and development on their platform.

Not vibe coding, just babysit and manual refactor. SwiftSyntax to slice source code in meaningful parts (done) and send it to openai ada (best for code embeddings). Then MCP for file manipulation. Just sharing the recipe, not selling anything, the moat is non-existent.