Share
Summarize with AI
MeDo Speaks SKILL.md: Custom Skills for Any AI Workflow
1. How a Markdown file became the common language of agent capabilities
SKILL.md became a standard because capabilities needed to travel between agents the way files travel between computers — and the ecosystem agreed on the container within about three months. A skill is a directory that must contain SKILL.md, whose frontmatter requires only name and description; license, compatibility, and metadata fields are optional, and supporting files live in optional scripts/, references/, and assets/ folders.
The load model is what made adoption painless. Agents see only a short metadata block for every installed skill up front, load the full body when a skill matches the task, and read reference files on demand — so you can install hundreds of skills without drowning the model's context, per the open agent-skills specification. Anthropic introduced the format in October 2025 and opened it as a vendor-neutral standard in December 2025; OpenAI, Microsoft, and Google's Gemini CLI followed within weeks, and a community survey counted more than 280,000 published skills by February 2026 across more than two dozen agents. Vercel's npx skills CLI became a common way to install a skill straight from a Git repository — write once, run in any supporting agent.
2. What Skills looks like inside MeDo
Inside MeDo, Skills is a first-class surface rather than a developer setting: browse official skills in the Skill Center, add your own custom skills, and use any of them in a conversation or while generating an app. For people who direct an AI builder in plain language — the vibe-coding workflow MeDo is built around — a skill removes the need to re-describe your workflow each time: you invoke one packaged capability that already knows the details.
Official skills come in two kinds. Prompt-based skills carry a predefined prompt that keeps the model focused on a type of task, such as translation or OCR. API-based skills connect to external services for real-time data or third-party capabilities, and they can be wired into an app so the product you generate uses the service too. Together they span AI, voice, image, payment, search, and office productivity; MeDo picks the right one from context, or you specify it yourself with @.
Credit-consuming official skills run against a daily free quota — one video-generation call, five image calls, and 100 calls for everything else, as of September 2026. The free allowance is used first, resets at 00:00, and does not carry over; once exhausted, skills that normally consume credits charge their standard credit price, which the Skill Center shows.
3. Custom Skills: bring your own SKILL.md
A custom skill starts with a file you may already own. Under Skills → My Skills → Create Skill, you upload a .zip skill package or a bare SKILL.md; MeDo parses it — usually within a minute — and generates the skill definition and configuration for you:
- Upload the
.zippackage orSKILL.mdfile. - Review the generated name, icon, and description, and add environment variables the skill needs, such as an API key.
- Create the skill and invoke it with
@to test it. - Check the Used Skills list, then reuse it across apps and tasks — or manage and delete it later in My Skills.
Two details shape what you can upload. The package must follow the ecosystem's standard structure: SKILL.md is required, scripts/, references/, and assets/ are optional, the file name is case-sensitive, and a .zip's root directory must match the lowercase-hyphen name in the frontmatter. The skill definition MeDo generates is derived from the package rather than hand-editable — if the result does not fit, you adjust the source files and re-upload, as the Custom Skills guide details.
MeDo does not yet offer a create-from-blank form for custom skills. If you have no package, you scaffold one conversationally: mention the @Skill Creation & Editing skill with an API description or documentation link, and MeDo produces a downloadable skill package you then import — or, for a single API, hand MeDo the API description directly in chat to integrate the service. The file becomes the artifact you keep and version, exactly as in the rest of the ecosystem.
4. Two faster ways to reach a skill: @ and in-chat configuration
You should not have to remember which skills exist mid-request, so MeDo surfaces them where you type. Since August, the toolbar below the chat box opens the Skills list in one tap, and typing @ in the conversation does the same — official and custom skills both appear, custom first.
September added configuration without leaving the thread: when a conversation needs a user or custom skill, the Skills configuration entry inside the conversation manages it on the spot. The changes are tracked in the MeDo changelog; together they complete the loop — import once, reach a skill in seconds, tune it mid-conversation.
5. The other direction: MeDo ships as a skill other agents install
The ecosystem story is not one-way. The MeDo App Builder skill is published on ClawHub, the skill registry for OpenClaw, a self-hosted open-source agent. An OpenClaw user installs it from the CLI — openclaw skills install @seiriosplus/medo-app-builder — sets a MEDO_API_KEY in the environment, and then talks to their agent in natural language to list apps, inspect projects, create new MeDo apps, iterate on them, and publish to a production URL.
The capability set matters less than what it proves. The same folder standard that lets MeDo parse a third-party package is what lets a third-party agent load MeDo — identical conventions, identical expectations about discovery and invocation. MeDo is not an island that imports the format; it is a node on the same network, publishing to the same registries its users pull from. That is the part no proprietary integration could replicate as cleanly.
6. Managed parsing, not blind installs: where the boundary sits
One distinction deserves to be explicit because "skills" covers two operating models that behave differently. When you upload a package to MeDo, the platform parses the definition, stores credentials through its own configuration, and runs the capability in the managed builder environment — the same trust posture as any MeDo feature. When a local agent downloads third-party skill packages and executes their scripts on your machine, the trust boundary sits on your computer instead.
The community survey cited earlier also notes that a meaningful share of published skills carry vulnerabilities, so the standing advice for local agents is to review a skill's source before running it, as you would any installed code. That is not a claim that managed platforms are risk-free, nor a critique of local agents — it is simply where each model puts the boundary. The files are identical; the environment that executes them is not.
Conclusion
Three releases over one summer — Custom Skills in June, @ invocation in August, in-chat configuration in September — add up to one capability: MeDo treats the open SKILL.md standard as its skill format, and the open ecosystem treats MeDo as a skill. Knowledge you already packaged for other agents is no longer stranded when you build an app, and the agents you run locally can reach the apps you build here. It pairs with the built-in SEO Agent, covered in its own post earlier this week — agent capabilities are becoming a product line of their own rather than a footnote to the builder.
The fastest way to feel the difference is small: take a SKILL.md you already use elsewhere, upload it under My Skills, and ask MeDo to apply it to a task or an app. Then install the MeDo App Builder skill into a self-hosted agent and see the same standard from the other side. Everything you need is on the MeDo features page.
FAQs
Frequently asked questions
Keep reading
Related articles
MeDo's AI SEO Agent: Audits, Fixes, and Publishing in One Place
Run a website SEO audit, fix detected issues, and publish — without leaving your builder. How MeDo's AI SEO Agent turns on-page SEO into a single workflow.
MeDo Components: AI-Generated UI Blocks for Any Builder
MeDo Components is live: type a prompt and get a production-ready React + Tailwind UI block — navbar, pricing table, hero — for any AI builder.
Built-in SEO Agents Compared: MeDo vs Lovable vs Replit
MeDo, Lovable, and Replit all shipped built-in SEO agents by 2026. Compare their audit loops, measurement tools, and SSR defaults to pick the right one.