My LLM workflow: - generate .md file of entire code base, minus noise - paste...
My LLM workflow:
- generate .md file of entire code base, minus noise
- paste into chat, ask bot to do something
- Bot gives me new main.py and I create a m.py file, paste in new code
- Use #Jetbrain's diff tool to manually merge
Self-replies
If I'd used Cursor/Junie etc, the LLM will erase comments and sometimes large sections of code. The bot has a sense of "does this *really* need to be here? no? delete it" and that just isn't safe.
It will add "conversational comments" ("You should do this next")
Also, I just need to see what's going on.
Anyhow, I've gone from using the diff tool once a week to *all* my time is in the diff tool. I wish there was an easier route to diffing than creating that dummy file, typing a file name, telling Jetbrains, 'I don't care if this is in git'
Anyhow this feels like a 3rd, underhyped workflow.
Other workflows:
- AI autocomplete. I actually don't do this anymore.
- Agentic (codex, etc). When I try, it is slow and I don't like how the bot misunderstands me and I waste $$ tokens on code I have to discard now.