Cursor
Cursor reads .agents/skills/ natively, plus .cursor/skills/ and compatibility reads of .claude/skills/ and .codex/skills/. No skills projection is written. The canonical directory is already a Cursor directory.
Cursor is enabled unconditionally alongside Claude Code, so the command reports it as enabled whether or not .cursor/ exists. Nothing is written for it either way.
The interesting part is instructions.
The mode split
Section titled “The mode split”Cursor reads AGENTS.md in Agent mode only. Chat and Composer read .cursorrules and .cursor/rules/*.mdc instead.
If that is right, a repository that consolidates everything into AGENTS.md keeps its instructions in Agent mode and loses them in Chat and Composer, two of three surfaces.
This claim is contested. It comes from third-party comparisons, not from Cursor’s own documentation. It is consequential enough to be worth verifying against Cursor’s documentation before you rely on it either way, and it is stated here as unconfirmed rather than as settled fact. See Sources & Confidence.
What not to do
Section titled “What not to do”- Do not generate
.cursor/rules/*.mdcfromAGENTS.md. Rules are canonical-only..mdcand.mdare not interchangeable, and path-scoping has noAGENTS.mdequivalent. - Do not delete or rewrite
.cursorrulesor.cursor/rules/**on the assumption thatAGENTS.mdcovers them. In Chat and Composer it does not. - Do not assume Cursor’s
.agents/skillsdiscovery recurses into nested subdirectories. That is untested.
The practical position: consolidate instructions into AGENTS.md, and leave the Cursor rules files alone rather than trading a confirmed loss for an unconfirmed gain.
Frontmatter
Section titled “Frontmatter”paths, disable-model-invocation, and legacy globs are Cursor-recognized. Other harnesses drop them, so keep load-bearing behavior in the Markdown body. See Portable Skills.