I Turned Claude Blog Into a Web Skill Vault
Listen to the community summary · 1 minute 17 seconds
Read the audio transcript · This is the original short summary, not a narration of the full article.
Hey homies. I turned Claude Blog into a skill vault you can use straight from the web.
The idea started with a simple problem. You install skills, collect more skills, update some and forget the others. Eventually, you're maintaining a small museum of instructions.
I wanted a different way to use that knowledge: give an agent a website address and a task, let it find the relevant instructions, and get to work. That's what I built with Web Skill MCP.
The idea: Keep the skills and their references in a Markdown-first website. Let agents read what they need for the current task, either through the web or a connected MCP client. Individual skill installation is optional; the tools needed to execute the task still matter.
watch the experiment
In this walkthrough, I give Claude the vault link and ask for a blog outline about which countries have the best tomatoes. Watch it find the outline skill and start researching.
A website that holds the workflow
The foundation is Markdown. The vault holds skill instructions, reference files and guidance on when to use them. You can browse it like a folder of notes, inspect a file, and follow its links to related material.
The release we tested on September 16, 2026 contained 32 skills in that explorer, with source versions traceable to the GitHub repository. The vault's source guide explains the import scope and the distinction between repository content and guidance written for the website.
That matters because a skill is more than a prompt with a clever title. It can include a workflow, templates, supporting references and scripts. The Agent Skills documentation describes a format built around a SKILL.md file, with supporting material loaded when needed.
The folder structure makes those relationships visible. When a writing skill points to a research reference, the agent needs a way to reach that reference. Otherwise, we've moved the first page of the instructions online and left the rest behind.
One vault.
Just what you need.
Select a workflow to explore the connection.
Outline: load the structure guidance, then plan the article using your brief.
Illustrative workflowGive your agent a task and the vault URL
The easiest way to try it is with an agent that can browse or fetch web pages. Give it the address and a specific outcome.
For example:
Use https://claude-blog.md/mcp to create a blog draft about
how a small team can document its weekly decisions.
Write for team leads who want a practical starting point.
Read the agent guide, choose the relevant skills, and load
the references needed for this task.
Research first, outline second, then write and review.
Cite sources you actually read. Tell me which checks you
couldn't perform. Keep the result as a draft.
Ask the agent to find the relevant workflow, read its instructions and apply them. Whether it completes those steps depends on its access and capabilities. You can ask it to name the files it used so you can check whether it followed the references.
For a smaller first test, ask for an outline. That gives you a concrete result to inspect before asking the agent to research and write a full article.
If your agent supports remote MCP, you can use Connect MCP in the vault. Once connected, the client can search for skills and load them through tools. The compatibility guide covers both routes and their requirements.
How this relates to MCP and WebMCP
I use Web Skill MCP as the name for this project. It combines a readable website with ways for agents to retrieve the same content.
There are three access routes:
| Route | How it works |
|---|---|
| Browser or HTTP | The agent reads the Markdown and follows its reference links. |
| Remote MCP | A connected client calls tools to search, describe and load skills. |
| Browser WebMCP | A supported browser integration calls the tools exposed by the page. |
Standard MCP connects AI applications to tools and context. Its official SDK documentation describes remote connections alongside other transports. Our endpoint uses that connection to deliver skill guidance.
Google's WebMCP introduction describes structured actions exposed by websites for browser agents. In our built-in browser test, the page's WebMCP interface successfully described and loaded the outline skill. That's a result for this browser and this page; it doesn't establish support in every agent.
The Markdown route keeps the content usable even when those integrations aren't available. An llms.txt index provides another entry point to relevant files. The agent still needs to find or receive the address and decide what fits the user's task.
No local skill install, but real tools still matter
Loading a writing skill doesn't magically connect your WordPress account.
The vault supplies instructions. Your agent supplies the capabilities to follow them. Live research needs web access, image generation needs an image tool, and publishing needs a configured destination and your permission.
Some installed workflows also rely on local scripts. The hosted writing skill declares that its blog_preflight.py validation script isn't bundled. Reading the workflow therefore doesn't mean the agent has run every check that an installed setup could run.
This is the distinction I want people to understand before trying it. You can use the hosted guidance without installing each skill. You still need to check whether your agent can complete the particular task.
The agent guide and trust guidance keep that use scoped to the user's request. These are instructions for the current job, not permission to change accounts, publish content or override the agent's existing rules. Citations and factual claims still need review.
What happened when we tried it
We used the live endpoint to retrieve the brief, outline and writing skills, then loaded the supporting research and delivery references. We researched the topic, created an outline and produced a draft.
One detail stood out: our broad search returned an ambiguous match. The agent had to inspect the descriptions and choose the right skills. Hosting the library makes it accessible; choosing the right workflow remains part of the work.
We also opened the vault in the built-in browser and called its WebMCP tools to describe and load blog-outline. That showed the browser route working for those retrieval actions, alongside the separate remote MCP requests.
The experiment notes retain the exact release, tool responses and checks. They remain in the local draft package, together with a browser walkthrough. We haven't measured universal compatibility, token savings or identical output quality against every installed setup.
- 01Research
Read sources.
Check the claims. - 02Outline
Choose the structure.
Plan the evidence. - 03Write
Apply the skill.
Keep citations clear. - 04Review
Test the output.
State any gaps.
Try one task
Open claude-blog.md/mcp, choose a task and ask your agent to use the relevant skill. Start with a brief, an outline or a rewrite.
Then inspect the result. Which instructions did it read? Did it follow the references? Did it say when a required tool was missing?
That's the experiment I wanted to share with the community: fewer individual installations, a readable home for the workflows, and the same responsibility to check the work.
Written by Daniel Agrici, creator of Claude Blog. Adapted from my community audio summary with AI assistance, using the hosted Claude Blog guidance.