This looks dope! gonna check it out

Looks great!

@Jonathan White Essentially showing proof that everyone is using the agents/claude.md file wrong. Usually we use it to ensure Claude follows our guide line/set parameters but it does work as straight forward as we think. Developer-written context files only improved success by about 4%, while LLM‑auto‑generated context files actually made performance about 3% worse and increased costs by 20%. Core advice: Do not dump repo overviews, architecture, and “everything important” into CLAUDE․md/AGENTS․md. If the info is already in the codebase, the model can usually find it; repeating it in context just bloats and distracts. Big context files are not great as they age quickly and go out of date, causing the agent to place files in wrong locations or follow old patterns. - What you should do instead - Let the agent explore the repo and rely on package.json, file structure, and tests; modern models are already good at this. - Focus on improving your codebase: clearer structure, better tests, type checks, and tools that give strong feedback, rather than over-engineering the agent prompt. -Use CLAUDE․md/AGENTS․md as a small, targeted “steering” file only when a model is consistently making a specific mistake (e.g., “always run type checks after changes”) Theo’s practical pattern - Start with no or almost-empty agent file; only add rules when you see repeated failure modes.​ - Sometimes “lie” in the file to push the agent in a productive direction (e.g., telling it the app is greenfield so it feels free to refactor schemas instead of over-focusing on data migrations).​ - Use the agent’s struggles as a signal to change the code, not to keep piling rules into CLAUDE.md​.

Will be there 👍