# LLMs
## Guard Rails
aka personal LLM veganism
### I will not use LLMs for
>I will not use LLMs for anything with a face
- Writing. Writing is useful as a thinking tool and outsourcing thought to a machine is to my hurt.
- [Code reviews](https://dd.slack.com/archives/CNF2V6Z4J/p1754925426990149?thread_ts=1754921217.509859&cid=CNF2V6Z4J) (slack link)
- Software systems are theory building and comprehending teammates code is critical to that process. Outsourcing that to a machine hurts my team and feels wrong.
- [Automated PRs](https://bauer.codes/post/2025/07/ai-pull-requests/)
### I will use LLMs for
- summarizing walls of text, especially walls of text others used AI to generate
- researching topics
- I've found LLMs freakishly good at taking an english language description of some problem and suggesting an obscure algorithm Knuth catalogued in TAOCP.
- Have to be careful though and verify everything. I once asked claude and chatgpt the same question about a drug interaction and they both gave opposite, equally persuasive answers. Seems ... bad.
- Pair programming. It's hit and miss, but on balance, Claude's better at boilerplate construction than I am copy-and-pasting from stack overflow. Just like SO, it's a mixed bag, but relatively straightforward to verify. I seldom accept its solutions as-is as I find its design sense verbose and mostly awful.