innernet · skills
what is a SKILL.md
the one file behind every ai skill, what goes in it, and why the description line does most of the work.
A SKILL.md is a markdown file that teaches an AI how to do one job. It sits in a folder named for the skill, it starts with a short block of metadata, and everything after that block is instructions written for a model to follow.
That is the whole format, and the plainness is deliberate. Anyone who can write a document can write a skill, any tool that can read a file can load one, and the same file works in more than one place.
the shape of the file
---
name: weekly-report
description: >
turn a week of raw notes into the team's weekly report. use when someone
asks for the weekly update, a status report, or "what happened this week".
---
# weekly report
1. group the notes by project, not by day.
2. lead each project with the one thing a reader must know.
3. ...Two things are required: a name and a description. The name is a short, lowercase, hyphenated handle. The description says what the skill does and when to use it, in the words that moment actually uses.
Below the metadata is ordinary markdown: steps, rules, examples, a checklist to run at the end. The folder can hold more files beside it, such as a script, a template or a long reference table, and the main file says when to open each.
how an ai decides to load it
This is the part that makes a skill different from a long prompt.
When a session starts, the AI does not read every skill in full. It reads the name and description of each one, which costs very little, and keeps the rest on the shelf. When your request matches a description, it opens that skill and reads the instructions. When the instructions point at a supporting file, it opens that file only if the job needs it.
So you can keep dozens of skills available without filling the conversation with them, and the description is the trigger. A vague description ("helps with reports") means the skill is rarely loaded when it should be. A specific one ("use when someone asks for the weekly update") means it is.
what goes in, and what doesn't
In: durable method. How to do a kind of work well, the order to do it in, the rules and the reasons for them, the mistakes that make the output wrong, and a checklist.
Out: the task in front of you (that's a prompt), facts about a person (that's memory), anything private if you plan to share it, and anything the code or the data already says, which will go stale in a document.
A good test: would this still be true next month, for someone who isn't you? If yes, it belongs in a skill.
a skill, a prompt, an instruction, a server
People mix these up because they all change what an AI does. They do different jobs.
| what it is | how long it lasts | when it applies | |
|---|---|---|---|
| prompt | what you type in one chat | that conversation | now |
| custom instructions | a standing note to one tool | until you change it | every conversation, in that tool |
| skill | a document teaching one job | until its author changes it | when the job comes up, in any tool that reads it |
| MCP server | a connection to a system | while it's connected | when the AI needs to act on or read that system |
A skill teaches how. An MCP server gives access. They are often used together: a skill for how to read a sales report, a server that fetches the report. MCP servers vs skills goes further into that split.
the fields worth adding
Beyond name and description, a few optional fields make a shared skill much more useful to the next person:
- when it was last changed, so a reader can tell a maintained skill from an abandoned one;
- who wrote it, and where it came from, so a reader knows whose judgement they're following;
- a version, so a team can tell which state they're using;
- a licence, so people know whether they may adapt it.
innernet's published skills carry all of these. The format innernet uses is itself a skill, the skill format, which you can read, or paste into your AI and ask it to check a skill against.
where to see real ones
The fastest way to understand the format is to read five good skills. The commons keeps skills from Anthropic, Vercel, Supabase, Cloudflare and individual makers, each readable as a page and fetchable as raw markdown by adding /raw to its link.
questions
what is a SKILL.md file?
A SKILL.md is a markdown file that teaches an AI agent how to do one specific job. It starts with a metadata block holding at least a name and a description of when to use it, followed by instructions written for the model. It usually lives in a folder named after the skill, alongside any scripts or reference files it needs.
what fields does a SKILL.md need?
A name and a description. The description should say both what the skill does and when to use it, because the AI reads the description to decide whether to load the rest. Author, version, last-updated date and licence are optional but make a shared skill trustworthy.
how is a skill different from a prompt?
A prompt is typed into one conversation and ends with it. A skill is saved, loaded automatically when a request matches its description, and reused across sessions and, where supported, across tools.
how is a skill different from an MCP server?
A skill teaches an AI how to do a job. An MCP server connects an AI to a system so it can read data or take actions there. A skill is a document; a server is running software.
do skills work outside claude?
The format is open and plain markdown, and several agents and editors besides Claude read skills. Where a tool does not read skill folders, a skill can still be handed over as a link or pasted in, because it is only a document.
innernet is user-owned memory across AI tools, reachable over MCP. The skills you keep live beside your projects and decisions, and every tool you connect reads the same copy. innernet.live
Checked 17 september 2026.
keep reading
one memory, every ai tool.
innernet keeps your projects, your decisions and your skills in one place, and every ai tool you connect reads the same thing. the skills people keep there are open to anyone.