One SKILL.md folder that Claude Code, Cursor and Codex all pick up (plus a flaky-test skill you can steal)

Claude, Cursor, and Codex each pick up skills from a single SKILL.md folder, making the process more efficient. This shared format allows the same folder to be used across all three tools. The skills solve the issue of long AGENTS.md files that get loaded into every session, even when they're not needed.

A skill is a folder with a SKILL.md file. When the task calls for it, the agent only sees the name and description of the skill and reads the body when necessary. This means that rules, such as how to review PRs, can be moved out of the always-on file and only cost context when needed.

The format for naming and structuring skills is standardized, as outlined in the Agent Skills spec. To use a skill, you can invoke it in the agent chat with a command like /skills or $ mention.

Here's an example skill called fix-flaky-test that can be copied and used right away. It provides a step-by-step process to diagnose and fix flaky tests, including reproducing the flake, identifying the source of nondeterminism, and fixing the cause instead of adding retries. The skill emphasizes not adding retries or skipping tests without a valid reason tied to the cause.

Some gotchas to be aware of include broken YAML that fails quietly in Claude Code, name clashes between project and bundled skills, and how Codex and Cursor handle the initial skills list. The author recommends committing the needed skills to the project repository and provides a link to purchase the Agent Skill Pack for those who prefer not to write them themselves.