Best AI Agent Skills for Keeping Docs Up to Date (2026)

Nobody's docs are wrong on the day they ship. They rot. An endpoint gets deprecated, a button moves, a diagram describes an architecture that was replaced two quarters ago, and the page keeps sitting there looking authoritative.

Most writing about AI and documentation is about producing the first draft, which was never the expensive part. The expensive part is the audit nobody schedules: reading 400 published pages to find the 30 that now lie.

That job is mechanical, tedious, and unbounded, which makes it the thing agent skills are best at. Below are the skills and MCP servers worth installing for it, with install commands from each project's repo and honest notes on trust tier.

The stack at a glance

Crawl your own docs and ask what's a lie

Firecrawl scrapes and extracts structured data across a site, 7,159 stars, MIT, community-reviewed:

claude mcp add firecrawl -- npx -y firecrawl-mcp

Point it at your own documentation. Then ask questions you'd never assign a person: which pages still reference the v2 endpoint, which pages mention a pricing tier that no longer exists, which code samples import a package we dropped.

This is a two-day manual job that becomes an afternoon, and the reason it's worth automating isn't speed. It's that a two-day manual job never gets approved, so it never happens, and the errors accumulate for years.

How you phrase the question decides whether this works. "Find outdated pages" returns noise, because the agent has no definition of outdated. "List every page containing the string v2/billing, with the surrounding sentence" returns something you can act on in five minutes. Give it the deprecated names, the retired plan tiers, the renamed products. The agent is doing retrieval at scale, not judgement, and it performs exactly as well as the list you hand it.

One caveat that matters for internal docs: Firecrawl's default plan sends pages to a hosted service. For anything not already public, check the self-hosting option or use a different approach.

Screenshots have a shorter half-life than prose

Every UI screenshot in your docs is a small liability on a timer. The button moves, the label changes, and the getting-started guide now shows a version nobody runs. Readers notice this faster than they notice a wrong sentence, and it costs more trust.

Playwright MCP is Microsoft's, Apache-2.0, 35,883 stars, audited:

claude mcp add playwright -- npx @playwright/mcp@latest

Script the path to each documented screen, point it at a running instance, and capture the images as part of a release. It's visual regression testing aimed at your docs instead of your test suite. There's also a Playwright skill at 2,985 stars if you'd rather have instructions than a server.

This pays back faster than anything else on the list and gets adopted least. An afternoon of setup removes a chore that otherwise recurs every release, forever.

The spec-completeness audit nobody produces

OpenAPI MCP lets an agent navigate a large spec without loading the whole thing into context, 900 stars, MIT, unreviewed:

claude mcp add openapi-mcp-server -- npx -y janwilmake/openapi-mcp-server

For a 12,000-line spec that's the difference between usable and not. Ask which endpoints have no description, which parameters have no example, where a response schema contradicts the prose page describing it.

Handing engineering a list of 60 undocumented parameters is one of the most useful things a writer can do, and almost nobody produces one, because reading a 12,000-line YAML file by hand is nobody's idea of a Tuesday. Our API design and OpenAPI skills roundup covers the engineering side.

Finding the pages everyone forgot

MCP Atlassian covers Confluence and Jira, 5,706 stars, MIT, community-reviewed:

claude mcp add mcp-atlassian -- npx -y sooperset/mcp-atlassian

The query worth running quarterly: every page in this space not updated since the v3 launch, who owns it, and whether anyone has viewed it this year. That produces a delete list, and a delete list is worth more than a backlog of rewrites. Most documentation problems are volume problems.

Notion MCP is official and audited at 4,578 stars:

claude mcp add notion -- npx -y @notionhq/mcp-server

And Obsidian Skills from kepano is the highest-starred item here at 44,310, teaching agents Markdown, Bases and JSON Canvas:

claude skill add kepano/obsidian-skills

If your drafts live in a local vault before reaching the official system, that's the one that fits. We went deeper in our Obsidian agent skills guide.

Diagrams: make them regenerable, not perfect

The reason architecture diagrams go stale is that they're expensive to change. Someone drew boxes in a canvas tool, nobody has the source file, and updating one arrow means redrawing the thing.

The fix is making diagrams cheap to regenerate. drawio-skill, 7,262 stars, produces .drawio files, which stay editable:

claude skill add Agents365-ai/drawio-skill

Its licence is listed as Unknown in our directory and it's unreviewed. Lower stakes than a database connector, but check the repo if your employer has a licence policy.

Mermaid MCP is the better answer for docs-as-code teams, because Mermaid source lives in the repo next to the prose and diffs like code. 233 stars, MIT, unreviewed:

claude mcp add mermaid-mcp-server -- npx -y peng-shawn/mermaid-mcp-server

It renders to PNG, which matters when your site generator doesn't render Mermaid client-side or your PDF export drops it. Excalidraw MCP at 2,256 stars gives you the hand-drawn look, which honestly signals "this is current thinking, not a contract."

For pulling reference material in to check your docs against, Markdownify MCP converts nearly anything to Markdown, 2,907 stars, MIT:

claude mcp add markdownify-mcp -- npx -y zcaceres/markdownify-mcp

Where this approach fails

An agent can tell you a page contradicts the current API. It can't tell you the page is misleading while being technically correct, which is the more common and more damaging failure.

It also generates false positives at a rate that will annoy you. A page mentioning the v2 endpoint may be a migration guide that's supposed to. Treat every output as a list of candidates for a human to check, and don't let anyone wire this into an automated deletion.

The honest framing: these skills change documentation maintenance from a job nobody does into a job someone does in an afternoon each quarter. That's a large improvement and it is not automation.

A quarterly routine worth stealing

Crawl the published site with Firecrawl and list pages referencing anything deprecated. Run the OpenAPI completeness check. Pull the stale-page report from Confluence or Notion. Regenerate screenshots for the top 20 most-visited pages.

Half a day, once a quarter. Compare that against how long it's been since anyone read your docs end to end.

Related reading: tools for technical diagrams and documentation, agent skills for technical writers for the authoring side, and code documentation tools. Our full list for writers covers the wider toolkit.

FAQ

How is this different from a broken link checker?

A link checker finds URLs that 404. This finds pages that resolve fine and say something untrue, which no linter catches because it requires knowing what your product currently does.

Do I need to be technical to run these?

The MCP servers need a terminal and Node.js. If you've run a static site generator locally, you're fine. The claude skill add items are easier.

Which one should I install first?

Firecrawl, pointed at your own public docs. It produces a finding within an hour, and a concrete list of wrong pages is the argument you need to get time allocated for the rest.

Are these safe for internal documentation?

Check trust tier and read the repo. Playwright, Notion, Atlassian and Markdownify are audited or community-reviewed. drawio-skill, Mermaid MCP and OpenAPI MCP are unreviewed. Firecrawl sends pages to a hosted service on its default plan, which rules it out for confidential material unless self-hosted.

Can I use these outside Claude Code?

The MCP servers work with any MCP client, including Cursor, Windsurf and Zed. Skills installed with claude skill add are Claude-specific.