I think we will end up having prompt tracking aligned to commit SHA, stored in a separate file in the repo. In the case of micro steps, one commit == one prompt. In general, x prompts to achieve one commit.

Meanwhile, GH/VS Code, serving the late majority, requires more control and checks. GH's greatest strength—market reach—has become its current limitation.

I agree quality is not fully there. Still, it's only a matter of time. For 2025, I envision a process that begins with LLM (provide me with a baseline), followed by human input to refine v0 of the code (either directly or through prompts), rinse and repeat. This would be quicker for 70% of the work.

They know that LLMs will become commoditised over time, and the diversification between commercial and OSS will become increasingly limited. Market dominance by API/models is not sustainable. The money is in workflows for the enterprise and becoming the de facto "personal assistant" for consumers.

I did the same a few years ago. Not perfect, but I agree it's a worthwhile alternative to Google.

Over the past 6-12 months, I've been using Perplexity more.

The Model Context Protocol(modelcontextprotocol.io/introduction) recently introduced by Anthropic is a step in that direction.