What is an Agent Skill?
How reusable skills package instructions, workflows, and operating knowledge.
An Agent Skill is a reusable package of instructions and workflow knowledge. It helps an Agent behave consistently across repeated tasks.
The problem Skills solve
Without a Skill, users repeat the same instructions over and over:
“Review this PR carefully."
"Focus on regressions."
"Do not give me generic feedback."
"Mention tests."
"Use file references.”
A Skill turns that repeated instruction into a reusable operating procedure.
The mental model
A Skill should answer three questions:
- When should the Agent use this?
- What steps should the Agent follow?
- What does good output look like?
If a Skill cannot answer those questions, it is probably a prompt snippet rather than a teaching artifact.
Skills are especially useful for judgment work
Skills shine when the Agent needs judgment, not just access. A browser MCP can open a page. A debugging Skill can tell the Agent how to investigate a bug without guessing. A translation Skill can define tone, terminology, and review checks.
The best setups combine both: a Skill defines the method, and tools provide the evidence.
Compatibility needs evidence
Skill ecosystems are not identical. Some use SKILL.md. Some use rules. Some use slash commands or project memory.
For public recommendations, agentk.it should only show compatibility when the source documents it or a reviewer has verified it.
A Skill is the method layer. It does not give the Agent a new external tool; it teaches the Agent a repeatable way to reason, inspect, decide, and produce output.
- The task requires repeatable judgment: code review, debugging, planning, translation, research synthesis, release checks, or security review.
- You want the Agent to follow a consistent method every time instead of improvising from a one-off prompt.
- The team has standards that should travel with the repository or workspace.
- The Agent needs live external data or actions; use MCP or another tool integration instead.
- The instruction is too vague to test, such as 'write better code' or 'be smarter'.
- The Skill claims compatibility with many Agents but does not document how each Agent loads it.
A useful Skill says when to use it. The trigger can be a task type, file pattern, workflow stage, or user intent.
Look for concrete steps: inspect sources, compare behavior, run checks, classify findings, or produce a specific output.
Good Skills describe what the answer should contain and what mistakes to avoid.
Check whether it is for Codex, Claude Code, Cursor, OpenClaw, or another Agent. Similar ideas are not the same as compatible formats.
Code review Skill
The Skill tells the Agent to focus on regressions, missing tests, risky migrations, and file-level findings before writing a summary.
Skill library
A curated Skill repository can be useful, but it needs review because large collections often mix strong workflows with vague prompt packs.
- Calling every prompt a Skill. A Skill should package a reusable method, not just a clever sentence.
- Assuming a Claude Skill automatically works in Codex, Cursor, OpenClaw, or Qwen Code.
- Publishing huge Skill collections without reviewing source quality, license, and compatibility.
Use the checklist on one candidate Skill before publishing it: trigger, method, quality standard, compatibility, and source trust.