Using a skill day to day
How to phrase a request, let the agent pick the skill or call it by name, combine several skills and adapt one to your project.
Updated on
An installed skill needs no configuration: you describe your task and the agent reads the skill if it fits. This guide covers how to phrase the request, when to name the skill, how to chain several, and how to adjust one to your project.
What the agent sees of a skill
At startup the agent loads only each skill's name and description. The full SKILL.md file is read only when the task matches the description. That is the Agent Skills design: many skills can stay installed without weighing down the conversation.
In practice, the description alone decides whether a skill triggers. Read it on the skill's page before you write your request, and reuse its words.
Let the agent choose
The simplest method is to describe the result you want in the terms of the description.
- Weak: "Look at my site."
- Better: "Audit my homepage: structure, headings, call to action."
If the skill does not trigger, rephrase with words closer to its description. Claude Code's documentation gives the same advice: check that the description contains the keywords you would naturally say, then rephrase. The causes are covered in Why a skill does not trigger.
Call it by name
When you want to be certain a skill is used, name it.
| Agent | Explicit call |
|---|---|
| Claude Code | Type a slash followed by the name at the start of the message, for example /landing-page-audit. Whatever follows the name is passed to the skill. |
| Codex | In the CLI or the IDE extension, type the dollar sign to mention a skill, or run /skills to pick one from the list. |
In Claude Code the name only works as a command when it opens the message. In the middle of a sentence ("go ahead and /deploy to staging") it gives Claude permission to use the skill but does not run it directly. To talk about a skill without allowing it, leave off the slash.
Other agents have their own syntax: see their page in the catalog, for example Claude Code or Codex.
Combine several skills
There are two ways, depending on the agent.
In one natural-language request. Split the task: "Write the pricing page with the copywriting skill, then have it reviewed with the conversion audit skill." The agent loads each skill when its step comes up.
By stacking them, in Claude Code. In the terminal, several names at the start of one message load several skills. The text after them is passed as the argument to each:
/write-tests /fix-issue 123
Claude Code expands the first skill and up to five more. Stacking stops at the first skill that runs in a subagent. With claude -p or the SDK, only the first skill is loaded.
Combine only skills with distinct roles. Two skills that give conflicting instructions on the same topic (two style guides, say) lead to unpredictable output: pick one.
Example requests
| Need | Request |
|---|---|
| Audit | "Audit the login page for accessibility and list the fixes in priority order." |
| Generation | "Build a dashboard component from the project's DESIGN.md." |
| Review | "Review my uncommitted changes and flag risks before I commit." |
| Chaining | "Write tests for this module, then fix whatever fails." |
A good request gives the subject (which page, which module), the criterion (what matters to you) and the output you want (a list, a patch, a summary).
Adapt a skill to your project
After installation a skill is a folder of files you own. To adjust it:
- Install it at project level, for example in
.claude/skills/for Claude Code. Commit the folder and the whole team gets it. Each agent's folders are listed in the skills catalog and on the agent's page. - Edit
SKILL.md: add your conventions (folder names, test commands, tone of voice) or remove steps that do not apply. Claude Code picks up changes during the session, without a restart. - Keep it short. Once loaded, a skill stays in the conversation context, so every line costs tokens on every turn. Claude Code's documentation recommends keeping
SKILL.mdunder 500 lines and moving detail into separate files. - Keep the original license and notice, and note your changes.
Two useful settings in Claude Code: disable-model-invocation: true in the frontmatter stops the agent from launching the skill on its own (handy for a deployment), and the text typed after the name can be reused in the skill through the $ARGUMENTS placeholder. Codex has the equivalent of the first with allow_implicit_invocation: false; explicit calls with the dollar sign still work.
When the same skill exists in several places, Claude Code applies a priority: a personal skill wins over the project one. Delete the copy you no longer want.
Going further
If this is your first install, start with Install your first resource. To decide between a skill and a plain prompt, read Resource or plain prompt?.
Sources
Frequently asked questions
Do I have to name the skill in my request?
No. An installed skill is picked up when your request matches its description. Naming it is useful when you want to be sure it is used.
Can I use several skills in one request?
Yes. Describe the task in steps, or name the skills. Claude Code also accepts several skill names at the start of a message typed in the terminal.
Can I edit an installed skill?
Yes, it is a folder of files in your project or your home directory. Edit the copy, keep the original license and notice, and note what you changed.