Glossary · Context
Repo-local, versioned instructions agents load by convention: Claude-style project memory (CLAUDE.md), portable agent bootstrap (AGENTS.md), and Cursor-style skills (typically SKILL.md under .cursor/skills/).
These names refer to files and folders you check in so agents inherit the same stack, commands, conventions, and guardrails on every run—versioned context, not one-off chat.
In Claude Code and similar Anthropic-oriented setups, CLAUDE.md (usually at the repo root) holds persistent project memory: how to build and test, directory layout, coding style, and “how we work here.” The tool loads it into context so sessions replay the same baseline instead of re-explaining the repository each time.
AGENTS.md is a portable filename several agent products use for the same class of content: who the agent is helping, safe commands, test entrypoints, and sensitive areas. Think of it as a vendor-neutral sibling to CLAUDE.md—same job, filename some teams standardize across IDEs and CLIs.
Cursor (and similar workflows) use skills: reusable guidance stored as directories under .cursor/skills/, each with a SKILL.md (singular) the UI surfaces as a “skill.” They are scoped playbooks you opt into per task, alongside—not instead of—project-wide files like CLAUDE.md or AGENTS.md. The glossary bundles them here because teams talk about “CLAUDE / agents / skills” as one instruction layer.
See also: Context engineering, Skills, System prompt, Repo rules.