If you haven't used claude code, I highly recommend you check it out. You may have used other AI tools in development, but being a cli with a refined "plan" step has revolutionized my development workflow. The fact that I can dispatch three prompts to run concurrently, and reliably trust the actions being taken feels incredible. I am spending less time thinking about code and more about the higher level direction of the code.
If you want claude code to be more reliable in its output, it can further be directed by using CLAUDE.md files. These files are included in the plan step and significantly improve the quality of its output. What makes a useful CLAUDE.md is similar to a well written style guide [1]. By providing a set of guidelines for what patterns to follow, the LLM can be more focused when building its plan. In a sense, a CLAUDE.md can be thought of as a "language semantics" for your codebase or directory for how to "compile" your thoughts into actions. Fortunately, you don't need to write your semantics in BNF [2], but the completeness of your semantics matter nonetheless. Without forethought, an LLM will happily come up with its own plan for your codebase.
I thought of and published https://claudemd.dev last night with the idea that sharing these files will improve the quality of code that we will be writing for this next abstraction of software development.
If you want to learn more about these files, I recommend reading this https://claudelog.com/mechanics/claude-md-supremacy.
[1] https://github.com/airbnb/javascript [2] https://en.wikipedia.org/wiki/Backus%E2%80%93Naur_form
