@banteg - this applies to python: - for types, just tell it to run
banteg✓@banteg
"omg claude wrote 100k loc in x hours" is not a flex you think it is. the real measure of agent quality would be when it can reduce the loc. impossible challenge for claude, and somewhat challenging for codex.
i had to add some instructions to keep loc in check, and it took focused effort to reduce it from 12k to 8.8k loc. can't say the code is as minimalist as i would write it, but we'll get there.
rusowsky@0xrusowsky
@banteg how do u prompt it so that it leverages the type system, uses good abstractions, and minimizes LOC?
banteg✓@banteg
2025-12-25this applies to python:
- for types, just tell it to run `ty check`.
- for keeping complexity/loc in check, run `uvx lizard src/` periodically and ask it if there is anything actionable.
- for good abstractions, it would usually take some time for you to get the feeling of how the codebase should be structured. don't worry about it too early. when it becomes big, just send it to gpt/claude/gemini and ask for architectural improvements. feed it back to agent for refactor. after a few rounds of this, send it again and ask what abstractions are superfluous and cut them down. i admit this is bit of a vibe advice, and you would get better results if you actually know what you need instead of trusting llm advice.