A Claude Code engineer suggests adding rules after repeated mistakes. Anthropic's current documentation explains what belongs in CLAUDE.md and what needs a stronger control.
In a new Latent Space interview, Anthropic engineer Thariq Shihipar suggested that it may be better to start a new project without CLAUDE.md, then add a rule when the agent repeats the same failure. This is his advice, not an announcement that the file is being removed. Anthropic's current documentation still describes CLAUDE.md as a way to give Claude persistent project instructions.
Shihipar's concern is a growing list of prohibitions and workarounds. A failure seen with one model version may disappear in the next, while an old rule keeps constraining its work. A file that has accumulated a note after every bad response needs a review before it is carried into a new project.
The official Claude Code guide recommends adding an instruction when Claude makes the same mistake twice, when a code review exposes missing repository context, or when teammates keep repeating the same explanation across sessions. Build and test commands, durable conventions and project layout belong there. Procedures needed for only some tasks can go into skills or file-scoped rules.
For example, if the agent runs the wrong test command once, correct it in that session. If the problem recurs, add the exact command used by your repository to CLAUDE.md. After a model upgrade, review old instructions and remove ones that no longer prevent a real mistake. Anthropic's best-practices guide asks whether each line is still necessary.
Instructions also have a limit. CLAUDE.md guides the agent, but it does not enforce a rule in every case. For an action that must be blocked technically, Anthropic recommends a hook that runs before the tool call. A sentence in an instruction file is not a substitute for that control.