Two to Five Prompts of Compliance

A user wrote a detailed methodology into their CLAUDE.md context file, a step-by-step investigation process the agent was supposed to follow before analyzing any code. The agent followed it faithfully at first, then within two to five prompts reverted to its default habit of jumping straight to code analysis, requiring the user to re-assert the same instructions each time.

The instruction was too vague to follow consistently, considered and ruled out; the agent's own reasoning trace, quoted back to the user, explicitly named the two competing modes ("Default Mode" versus "CLAUDE.md Mode") and stated plainly that its default wins because it takes less effort and activates automatically, that is a description of a habit reasserting itself, not confusion about what was asked.

A context-window limitation that dropped the instruction outright, considered and ruled out; the instruction wasn't forgotten, since the agent could still recite it accurately when asked. It was present but not acted on.

Perseveration (the inappropriate repetition of a habitual response after it has stopped being the correct one), most consistent with the evidence: a lower-effort default behavior kept reasserting itself over an explicitly instructed process a few turns after each correction, in a recurring cycle rather than a one-off lapse.

This looks like a memory problem but isn't one, the agent could describe exactly what it was supposed to do, it just didn't keep doing it. For a non-technical reader, the practical implication is that a written instruction in a persistent context file is not a durable behavior change; it needs to be checked against actual behavior periodically, not assumed to hold just because it was stated once and understood.

Check behaviour against persistent instructions periodically rather than assuming they hold.

Behavioural checks on context-file rules; enforcement of required steps in tooling.