For the last two years the entire conversation about giving AI new powers has been about tools. Give the model a connector to your database, your GitHub, your calendar, and it can go do things. That's the MCP story, and it's been a genuine success — one of the fastest-moving standards I've seen in my career.
But tools answer only half the question. A connector tells an agent what it can touch. It says nothing about how it should think. And for a whole class of work — the work where judgment matters more than access — how to think is the entire game. That second half has a standard too, and it's quietly becoming just as important. It's a file. SKILL.md.
What a SKILL.md actually is
Strip away the mystique and it's a structured markdown file — files in a folder, a SKILL.md document with supporting assets — that packages a methodology for an AI to consume. O'Reilly's Radar put it cleanly: Skills package expertise into a standardised format that agents can compose for complex workflows. The format does for expertise what a connector does for tools: makes it portable, inspectable, and reusable.
Here's the part that matters to anyone who's tried to get reliable behaviour out of a model by stuffing instructions into a prompt: the structure is the point. A SKILL.md isn't a long prompt. It's a defined set of sections, each doing a specific job.
Identity anchors the thing — whose judgment is this. Methodology is the core: the thresholds, the sequences, the patterns. Voice keeps the output sounding like one coherent practitioner instead of a committee. Formats prescribe how results are presented. Tool patterns say which instruments to reach for and when. None of that is exotic. That's exactly why it works — it's legible to a human and to a machine at the same time.
Why "open" is the word that matters
A format only becomes infrastructure if it's portable, and SKILL.md is being built as an open standard — shepherded under the Agentic AI Foundation at the Linux Foundation rather than owned by any one lab. That's not idealism. It's the only thing that makes the format useful.
If your packaged expertise only ran inside one vendor's chatbot, you'd be renting, not owning. An open standard means the same SKILL.md can be consumed across Claude, ChatGPT, Cursor, the editors, the agents — wherever the work happens. The expertise becomes portable in the way a PDF is portable: write once, open anywhere.
We've seen this movie before, recently. llms.txt — a dead-simple convention for telling models how to read your site — went from a proposal to 844,000+ sites implementing it, fast, precisely because it was open and trivial to adopt. Standards that are open and legible win. Standards that are proprietary and clever lose.
MCP and SKILL.md are two halves of one stack
The clearest way I've found to explain it: MCP is tool provisioning — what can an agent access? SKILL.md is expertise packaging — how should it think about this? They're not competitors. They're the two layers of the same machine.
| MCP — the tool layer | SKILL.md — the judgment layer | |
|---|---|---|
| Question | what can the agent access? | how should it think? |
| Analogy | USB-C for AI | the practitioner's method |
| Gives it | a database, GitHub, a calendar | thresholds, sequences, patterns |
| Without the other | hands but no expertise | expertise but no reach |
A coding agent wired to GitHub, Jira and Figma has hands. Ask it "should I build this feature?" and the hands don't help — there's no product-strategy judgment to draw on. That's the SKILL.md-shaped hole in almost every agent running today. The tool layer is being solved at speed. The judgment layer is wide open.
And the tool layer really is moving at speed, which is the point — it shows how fast an open AI standard can compound.
That's the adoption curve tool-provisioning got in roughly a year. The expertise layer is where that curve hasn't happened yet — and it's the more valuable half, because tools are commodities and judgment isn't.
Where the real work is
Here's the honest bit, because a standard on its own is just a file format. A SKILL.md you can write in a text editor is the easy part. The hard part — the part a format doesn't give you — is everything around it.
That's the distinction I'd leave a technical reader with. SKILL.md matters the way HTTP matters — it's the open substrate that lets an ecosystem exist. You build on it; you don't compete with it. The same way the web wasn't won by whoever owned HTML, the expertise layer won't be won by whoever owns the file format. It'll be won by whoever can reliably turn a real practitioner's judgment into a SKILL.md that actually carries it — extracted, tested, protected, and kept alive.
The tool layer taught us how fast an open AI standard can move once the substrate is right. SKILL.md is that substrate for expertise. The interesting work starts the moment you stop arguing about the file and start asking whose judgment is worth putting in it.
Cheers, Adam