When teams adopt AI coding agents, the instinct is to add more context — specs, instruction files, documentation — whenever the agent makes a mistake. But past a certain point, more context degrades agent performance. Agents conflate documents, over-anchor on stale details, and walk confidently in the wrong direction. The core issue is conflating two kinds of truth: how the software currently is (best expressed through code and tests) versus how you want it to be (goals, direction, architectural constraints). Practical recommendations include cutting standing instructions, making context pull-in situational rather than automatic, enforcing architectural direction through folder/module structure rather than prose, reserving specs for genuinely complex subsystems like billing or gateways, and deciding upfront whether a spec is written for humans or agents. Some reference points remain valuable for verification, but the goal is the right context pulled in deliberately — not everything all at once.
