Create a Claude Skill That Triggers

Create a Claude Skill by giving a repeatable task a clear description and a procedure in SKILL.md, then installing it where you intend to run it. Use conversation to discover the rules or author the file directly when those rules are already settled.

Choose how to draft the procedure

Use a conversation when the workflow lives in someone's head. Give Claude a representative input and an approved output, describe where a reviewer usually intervenes, and ask Claude to draft both the procedure and the questions that remain open. Resolve those questions with the workflow owner before turning the draft into a skill.

Author the file directly when a checklist, template, and exception rules already exist. Create a skill folder with SKILL.md at its root. Start the file with YAML frontmatter containing name and description. Follow it with the expected inputs, the steps, and the response for missing information. Write the description so it identifies the task and when to use it. Add a supporting template if the procedure refers to one.

Test paraphrases and near misses separately. A broad description can increase recall while silently triggering the skill on unrelated requests.

Agent Skills - Claude Platform Docs

Install it in Claude for chat

For personal use in Claude, package the skill as a ZIP, confirm that code execution is enabled, and upload it in Customize > Skills. If you're on an Enterprise account and the controls are missing, ask the owner to check that Code execution and file creation and Skills are both enabled in the Policy tab under Organization settings > Plugins & skills.

Check that the uploaded skill appears and is enabled. Open a fresh chat, provide a sample input, and make a normal request that matches the description. Check for a distinctive output requirement, such as a section that lists missing inputs. If a good output could have been produced without the skill, use an exception case that requires its specific procedure.

Keep direct invocations in a separate test set. They are useful for checking the procedure, but they do not measure whether discovery chose the right skill.

Use skills in Claude | Claude Help Center

Run the same checks in Claude Code

For a development workflow, create SKILL.md in a skill folder in a location Claude Code discovers. Use Claude Code's direct invocation by skill name to check that it can find the file and follow the procedure. Then make the ordinary request in a fresh session to check whether Claude selects the skill on its own.

In an illustrative test of a skill called release-summary, supply notes with two user-facing changes, one internal refactor, and no release date. The expected result lists the two changes, leaves out the refactor, and flags the missing date. Ask for an internal code review as a control. If that request receives a release summary, narrow the description and repeat both requests.

Extend Claude with skills - Claude Code Docs

Diagnose the first point of failure

When Claude Code cannot invoke the skill directly, first check the actual filename, the installation location, and the frontmatter. When direct invocation succeeds but an ordinary request misses, compare the request with the skill description and add the vocabulary users reach for naturally. When the skill runs but invents a missing value, tighten the exception instruction and add an example that shows the expected response.

If the destination is an application calling the Claude API, configure that integration explicitly. The Messages API takes the skill type and skill ID in container.skills and requires the code execution tool. Use the same sample inputs to verify the application path and record its results independently of the interactive test.

Agent Skills - Claude Platform Docs

Sources