AI Agent Skills Best Practices
-
Keep it simple and stupid (KISS), think of it as a “mini-program” that do one or few things well
-
Use role framing: a brief description (one or two sentences) of the role that the agent will have when using the skill.
-
Define clear input signal & output format: always sepcify what data the agent expects to get as input and what output format it should be
-
Add examples
-
Add decision rules: agents may find ways to do something without explicit guidance, but that approach is not necessarily the best one. That’s why it’s better to tell how to think in specific situations.
-
Encode domain knowledge explicitly: since the agent will use the available data on the projects (i.e. source code, files, markdown files) to create a context or to find informations. If it doesn’t have some specific details, it will make assumptions that probably will not be what you or your client are looking for
-
Explicitly tell contraints and what not to do
-
https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices
-
https://uxplanet.org/7-rules-for-creating-an-effective-claude-code-skill-2d81f61fc7cd