What a skill is in Claude Code
A skill is a folder with a SKILL.md file. At the top of the file there are two fields: a name and a description that says what the skill does and when to use it. Below comes the instruction itself in plain text: steps, rules, an example of a good result. The folder can also hold reference files and scripts.
According to Anthropic's documentation, personal skills live in ~/.claude/skills/ and work in every project, project skills live in .claude/skills/ inside it.
- Name and description
One or two lines: what the skill does and when it is needed. The agent uses them to decide whether to take the skill.
- Instructions
Steps and rules in plain text: the order of work, the tone, what to check before sending.
- Reference files
Templates, examples, a price list or a procedure. The agent opens them only when it gets to them.
- Scripts
Programs the agent runs as a whole, for example to calculate or export data.
- Manual call
Any skill can be started with
/skill-namewithout waiting for the agent to pick it up.
How the agent decides which skill to use
In a regular session the agent sees only the names and descriptions of skills. When a task matches a description, it opens the skill and reads the full instruction, loading long templates and references as the work goes. That is why you can keep many skills: until a skill is needed, it takes almost no room in the model's context.
Hence the main rule: the description says in which situation the skill is needed. A description like "Reply to a customer review on a marketplace: tone, length, what not to promise" will be applied at the right moment. From the single word "Reviews" the agent can hardly tell which task it is for.
Why a skill beats explaining every time
- The same way of working. The instruction is the same, so this Monday's report is built the same way as last Monday's.
- Short requests. You write "make the weekly report", and the agent takes the format, sources and checks from the skill.
- Fixes add up. The agent made a mistake, you added one line to the skill, and next time it takes that line into account.
- A skill can be shared. The same file works for a colleague or in a second agent.
A skill and the agent's memory solve different problems. Memory holds facts: who your clients are, what was agreed, what your prices are. A skill records how to work on a task. How the agent remembers facts is covered in How to set up an AI agent.
Skill examples for business
| What's inside | When the agent uses it | |
|---|---|---|
| Review reply | Tone, length, how to answer a complaint, what not to promise | A new review arrived on a marketplace or in maps |
| Weekly report | Where to get the numbers, section order, what to compare with | Monday morning or on "weekly report" |
| Call notes | Decisions, tasks, deadlines and who owns what | After a call recording is transcribed |
| Commercial offer | A template, current prices, what to ask the client | A client asked for an offer or a quote |
| Channel post | Your voice, length, banned words, how to place a link | You ask for an announcement or a post |
| Email triage | Which emails are urgent, what to archive, what to forward to you | The morning inbox review |
It is easiest to start with one task you do at least once a week and explain the same way every time. More small business scenarios are in Examples of AI agents.
How to write your own skill
- Pick a recurring task and do it once together with the agent, correcting it along the way.
- Ask the agent to write down how you did it as a skill: steps, rules, an example of a good result.
- Check the description. It should say when the skill is needed, in the words you usually use to set the task.
- Move long templates and references into separate files next to it and link them from the instruction. Anthropic recommends keeping
SKILL.mditself under 500 lines. - When the agent makes a mistake, ask it to add a rule to the skill so the mistake does not repeat.
Skills and plugins
A Claude Code plugin is a bundle installed as one package: several skills, subagents, hooks and connections to outside services (MCP servers). Plugins come from marketplace catalogs via the /plugin command. A single skill works without a plugin: just put its folder in ~/.claude/skills/.
A plugin runs code with your user's permissions, so Anthropic recommends reviewing it before installing. The same goes for skills with scripts from other people: look at what is inside before installing them.
Skills in Codex
Codex by OpenAI works with the same open format: a folder with SKILL.md, a name and a description at the top. According to OpenAI's documentation, skills live in .agents/skills, you call them manually with /skills or $, and when a task matches the description Codex takes the skill by itself. So a skill written for one engine moves to the other with small edits: the folder changes, the instruction text stays the same. How else the engines differ is covered in Claude Code or Codex.
Anthropic's and OpenAI's documentation was checked on 2 November 2026. Formats and commands change, so check the service's help before you start.
Skills in NekoAgent
In NekoAgent skills live on your server next to the agent's memory and stay in place when you switch engines. You can ask the agent to write a new skill right in the chat: it works with files on the server.
Clients with a support subscription get the Skills library: ready-made skills for Zoom, calls, marketplaces, a coding agent and more. Iris installs them by code, which usually takes a couple of minutes.
We do the first setup together on a call with a curator: we connect your server, engine and Telegram, which takes about two hours.
We'll set up an agent you can teach with skills
From 24,900 RUB one-off per setup (Egoist plan). The server and the AI model subscription are paid separately. 30 days of support are included in the plan. The skills library is available with an active support subscription.
Key questions
What are skills in Claude Code?
A skill is a folder with a SKILL.md file: a name, a description of when to use it and instructions for the task. Claude Code sees the descriptions of all skills and opens the right one by itself when a task matches. You can also call a skill with /skill-name.
Where are Claude Code skills stored?
Personal skills are in ~/.claude/skills/ and work in every project on that computer or server. Project skills are in .claude/skills/ inside the project. Plugin skills are installed with the plugin.
How is a skill different from a plugin?
A skill is an instruction for one task. A plugin is a package that can hold several skills, subagents, hooks and MCP servers; it is installed from a marketplace with /plugin.
Do skills use up subscription limits?
Until a skill is needed, only its name and description sit in the model's context, which is small. The full text loads when the agent takes the skill and counts toward the limit like any request text.
Do skills work in Codex?
Yes. Codex supports the same open SKILL.md format, with skills stored in .agents/skills. A skill written for Claude Code moves to Codex with small edits.
