Hacker News Reader

How about leveraging llms to produce deterministic codemods? You can then iterate on this by running the codemods and using other deterministic guardrails, feeding the results back into the llm to improve the codemods?

Yeah exactly this is right. In the same way that we don't get an LLM to write the HTML for each request, but we do get an LLM to write the code to make the HTML

The framing in TFA is a dangerous way to think. Codemods are whatever, I could take em or leave em.

What you want isn't skill files for LLMs, though. Just write docs for humans. Write a migration guide, for humans.

It's going to take us a while to realize they should be the same things, skill files and docs.

Sorry, what is codemod here?

I just googled it a second ago https://martinfowler.com/articles/codemods-api-refactoring.h...

A tool to do automatic migrations eg react often publ

I think the two complement each other perfectly and will continue to do so. I keep writing AGENTS.md files for soft rules and custom linter rules for hard ones which IMO is the best of both worlds.