Hook

Their other posts in the index, biggest breakout first.
They cut Claude code system prompt by 80% and you should be doing the same thing to your Claude.MD files. The research shows that with newer models, less context can often yield better results cause you're allowing the model to make its own decisions about what context it loads instead of giving prescriptive rules. This applies just as much to your Claude.md file as it does to the Claude code system prompt. I guarantee your Claude.md file has too much information and is too long for these newer models. Dozens of independent research papers also back this up and certain benchmarks are actually able to take models like Sonnet and improve them by so many points just with context optimization that they perform the same as a model like Opus in the same harness with different context. There's no better time than now with models like Fable and Opus 5 available in Claude Code to optimize all level and project level Claude.md files to have reduced context and extract some of that context into skills instead so that agent can progressively disclose the information to itself when necessary. Link in bio for more sauce.