Claude Code keeps ignoring your CLAUDE.md
It is not you. Instruction compliance decays as a session runs, and resets after compaction. The fix is a mechanism, not a longer file.
Flowy indexes 2,590 Claude Code plugins. 58,811 components ship inside them: 39,896 skills, 9,239 commands, 8,990 agents and 686 MCP servers. 104 of those plugins fire themselves.Counted
Last updated
Why it happens
Your CLAUDE.md is read once, at the start. As the conversation gets longer, the agent weighs those early instructions less and less, and once the context is compacted into a summary, the rules are carried with even less force. The result is familiar: it followed your conventions for a while, then quietly drifted.
Writing a longer or stricter file does not fix it, because a longer file is still read once and still decays. What changes the outcome is re-firing the instruction, not writing more of it.
The fix: re-fire the rules
Re-assert the important rules on an interval and again after compaction, instead of trusting a single read. A hook can re-inject your routing and re-invoke the right skill each prompt, so the instruction stays live for the whole session.
This is not a workaround. Anthropic’s own Claude Code docs recommend hooks and re-invoking skills after compaction for exactly this problem. That is the mechanism Flowy implements.
Use hooks to enforce behavior deterministically. If the skill is large or you invoked several others after it, re-invoke it after compaction to restore the full content.
How Flowy keeps your rules firing
A Flow ships a FLOW.md router plus a hook. The hook re-asserts the routing on each prompt and after compaction, so the right skill keeps firing instead of drifting. Free and open source. You read every skill before you install.
Install the Flowy plugin
> /plugin marketplace add flowy-sh/flowy-core> /plugin install flowy-core> /plugin install flowy-superpowers> /flowy-superpowers:activate
FAQ
Why does Claude Code ignore my CLAUDE.md?
Instruction compliance decays as a session gets longer, and it resets after context compaction. Your CLAUDE.md is read once at the start; by the time the conversation is long or has been compacted, the agent is working from a summary that no longer carries your rules with the same weight.
How do I make Claude Code follow my instructions?
Re-assert the important rules on an interval and again after compaction, instead of relying on a single read at the start. A hook that re-injects your routing and re-invokes the right skill each time keeps the instruction live. Anthropic’s own Claude Code docs recommend hooks and re-invoking after compaction for exactly this reason.
Does a bigger or stricter CLAUDE.md fix it?
Not reliably. A longer file still gets read once and still decays. What changes the outcome is the mechanism: re-firing the instruction, not writing more of it.
How does Flowy help?
A Flow ships a FLOW.md router plus a hook. The hook re-asserts the routing on each prompt and after compaction, so the right skill keeps firing across a long session instead of drifting. It is free and open source, and you read every skill before you install.