# AI Instruction Drift AI instruction drift is a specific form of [[Context Drift]] that occurs when the instructions given to an AI agent (system prompts, CLAUDE.md files, skill definitions, rules) gradually diverge from what the user actually wants or needs. It happens through two mechanisms: **Accumulation without review.** Instructions pile up over time. Each addition makes sense in isolation, but the aggregate becomes contradictory or bloated. Rule 12 might conflict with rule 47. A skill added six months ago might encode assumptions that are no longer true. The [[Context Bloat]] compounds the drift. **Implicit evolution of intent.** The user's needs, preferences, and workflows change, but the instructions don't get updated to match. The AI keeps following the old playbook. This is particularly insidious because the user might not notice for a while; the AI still produces reasonable output, just not optimally aligned with current intent. Instruction drift differs from broader [[Context Drift]] in that it specifically affects the behavioral contract between user and AI. Drifted knowledge context might produce slightly less relevant answers. Drifted instructions produce systematically wrong behavior. Mitigation requires periodic instruction audits as part of [[Context Hygiene]]: reading through all active instructions, removing obsolete ones, resolving contradictions, and checking alignment with current workflows. [[Context Anchoring]] helps by making the reasoning behind instructions explicit, so you can evaluate whether the reasoning still holds even when the instruction itself looks reasonable. ## References - ## Related - [[Context Drift]] - [[Context Bloat]] - [[AI Context Rot]] - [[Context Hygiene]] - [[Context Anchoring]] - [[Context Engineering]] - [[Claude Code Memory]] - [[AI Agent Skills]] - [[Levels of AI Context Management]] - [[Configuration Drift]] - [[Obsidian Starter Kit - Tutorial - Managing AI sessions]] - Operational guidance to prevent instruction drift in OSK sessions