AI assistants don't lose track of your rules. They let newer context outweigh older instructions, which looks identical to forgetting but has a different fix.
The Details:
Every instruction in a session competes for attention rather than stacking cleanly. A constraint you stated in turn three sits against forty turns of error logs, stack traces, and new requirements, and the accumulated weight of recent context simply presses harder on the model's next move. Phrasing like "prefer not to touch X" reads as negotiable; a concrete trigger like "stop and ask if the change touches X" does not.
A bare prohibition is fragile, but a rule with a reason attached gives the model something to reason from. Telling it why a module is off limits, not just that it is, lets it apply the intent to situations you never explicitly listed.
Rules written only inside the conversation decay as the conversation grows. The fix is giving important constraints a structural home, like a short project instruction file with if-then conditions tied to actual file paths, each with a one-line reason. A twelve-page style guide fails because every rule inside it competes with a dozen neighbors for the same sliver of attention; a half-page of checkable triggers does not.
The strongest fix skips prompts entirely. File permissions, revoked database credentials, pre-commit hooks, and CI gates sit below the model's judgment and can't be argued around. The useful test for any rule is asking what's the lowest layer that can enforce it, then pushing it there, reserving prompts for the genuinely judgment-dependent cases.
Bottom Line: The rule was never forgotten; it was always just a suggestion competing for attention, so stop asking the model to remember and start making the dangerous path structurally harder than the safe one.
Enjoy this article?
Listen to the Claude Code Conversations radio show or join the community.
