CLAUDE.md
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
CLAUDE.md is a configuration/instruction file used to provide persistent context and guidance to AI coding agents — particularly Claude Code — governing how they should behave, what conventions to follow, and what constraints to respect within a given project or workflow.
Role in Agent Engineering
Ryan Lopopolo (OpenAI) notes that well-crafted instruction files of this type are a meaningful lever for shaping agent behavior: teams "can write good agents.mmd files" as part of disciplined agent engineering practice. → The implication is that the quality of these files directly influences execution quality downstream.
Risk of Unreviewd Plan Encoding
Lopopolo raises a specific failure mode tied to how CLAUDE.md-style instructions interact with agent planning: if an agent generates a plan and a human approves it without reading it, the approved plan effectively encodes instructions — potentially bad ones — that the agent will then follow during execution. As he puts it, "if you do use a plan and you approve it without reading it at all, you're actually encoding a bunch of instructions that you don't necessarily want followed." → This positions mandatory human review of agent plans as a prerequisite for CLAUDE.md-style guidance to function as intended, rather than as a source of unintended or poorly-considered behavioral constraints.
The practical implication is that CLAUDE.md is only as good as the attention paid to the full pipeline of instructions that flow into an agent — both the static file itself and any dynamically generated plans that supplement it.