My LLM workflow - If conceptually difficult, ask the bot to write a PEP-like or...
My LLM workflow
- If conceptually difficult, ask the bot to write a PEP-like or other design document
- Ask the bot to write code
- Carefully integrate it, run old school linter, unit tests, etc
- Try it out in real life
- Run code thru condenser/markdownifier so I can paste the whole context into a large context model
- Work with one bounded feature at at time
The bot almost *never* foresees problems I have with the design when used in real life scenarios.
Self-replies
Effort saved: none! I get much more ambitious. Everything has pluggy and uses all the argparse features
Catastrophic errors that were my fault vs bots - about equal. The bot making a catastrophic error is not that big of a problem, I can see it and tell it is off track.
Bot finding mistakes - it finds mistakes at about the same rate as unit tests & creates the same sort of design pressure. If the bot doesn't get where you're going with the code, you gave it a bad/fuzzy design goal.