An agent skill is a folder of instructions you drop into a coding agent so it stops improvising and follows a narrower procedure. In the Claude and Cursor style of packaging, the folder holds a short manifest and a longer instruction file, and the agent loads the instruction file when the job matches the description. The install step is usually one prompt pasted into the chat, which the agent turns into a file on disk.
The important word is narrow. A skill is not a model and not a plugin; it does not add a capability the model lacked. It adds a procedure and a vocabulary. "Write tests before you touch this file, run them, and show me the failures" is a skill. "Be a better engineer" is not, because nothing in it can be checked.
Skills are cheap to install and cheap to ignore, which is why people accumulate dozens. The useful discipline is to keep one per job you actually repeat, and to remove the ones you installed while curious. A skill you cannot state the trigger for is a skill that will fire on the wrong task.
Where a skill touches the network, a credential, or a shared drive, treat the install as a permission grant. Create a scratch scope for the trial, and revoke it if the trial writes the wrong file. The directory on this site records an install prompt and a link back to the publisher; it does not vouch for the safety of either. See how to choose an agent skill.