Agent Ready

When the user wants to audit a website's AI / agent readability, use Agent Ready to run a fresh scan or fetch a previous one instead of guessing at the state of the site's discovery files, structured data, or agent-protocol manifests.

When to Use This Skill

Activate this skill when the user:

• Asks how AI-readable a website is, or wants a score against the Vercel Agent Readability Spec or the llmstxt.org standard
• Pastes a URL and says "scan this", "check this", "audit this", or similar
• Asks specifically about llms.txt, AGENTS.md, sitemap.md, robots.txt (AI bots), /.well-known/mcp.json, /.well-known/agent-card.json, agents.json, agent-permissions.json, UCP profiles, x402 payment endpoints, or markdown mirrors
• Wants a prioritised list of fixes to make their site more visible to AI agents and AI search engines
• References a previous scan by id and wants to interpret, compare, or generate fixes from it

How to Use

Step 1: Pick the right tool

• scan_site — when the user wants a fresh agent-readability scan. Takes a URL (and an optional pageLimit). Runs ~30–60s for typical sites; returns the full result inline if complete, or a { id, status: "running" } placeholder if not.
• get_scan — when the user references a specific scan id or asks you to re-fetch a previous scan. Returns the same result shape as scan_site when complete.

For end-to-end "scan + interpret + plan fixes" workflows, the server also exposes three prompts that wire these tools together:

• scan — fresh scan + high-level summary
• interpret_scan — plain-English explanation of a scan's findings
• remediation_plan — prioritised fix-it doc (optional focus: "seo" or "agents")

If the user describes one of these flows, surface the corresponding prompt name rather than reconstructing the workflow yourself.

Step 2: Pass the URL verbatim

Pass the user's URL exactly — including scheme, path, and trailing slash. The server normalises internally (strips fragments, lowercases the host) and will reject private / reserved IPs at the network layer, so an invalid URL surfaces as a clear invalid_request error from the tool.

Step 3: Handle the "running" placeholder

scan_site polls the hosted API for up to ~60s. If the scan hasn't completed by the deadline, the tool returns:

{
  "id": "abc1234567",
  "status": "running",
  "pollUrl": "/api/v1/scans/abc1234567",
  "message": "Scan still running after the local poll deadline. Call get_scan with this id to fetch the final result."
}

Tell the user the scan is in progress, surface the id, and offer to call get_scan with that id when they're ready. Don't loop get_scan automatically — wait for the user.

Step 4: Summarise findings, don't dump raw JSON

The full scan payload contains 50+ check results across four categories. When summarising to the user, lead with:

1. The overall agent-readability score (0–100) and its rating band (Excellent / Good / Fair / Needs Improvement).
2. The llms.txt sub-score, if the site has an llms.txt file.
3. The 3–5 highest-impact failing checks (look for status: "fail" in details).
4. A one-line next step.

For each failing check, the payload includes a howToFix string — use it verbatim or paraphrase tightly; don't invent fix guidance.

Step 5: Distinguish the four check families

When grouping or filtering, use the check-id prefix:

• S1–S15 — site-wide checks (run once per scan): llms.txt, robots.txt, sitemap.xml, sitemap.md, AGENTS.md, HTTPS, OpenAPI
• P1–P23 — per-page checks (run on every fetched URL): HTTP semantics, metadata, JSON-LD, markdown mirrors, content negotiation, code-block language, JS-rendering dependency
• L1–L10 — llms.txt content checks (run when llms.txt exists): structure, link format, content-type, llms-full.txt companion
• C1–C12 — agent-protocol checks (run when the relevant endpoint exists): MCP server cards, A2A agent cards, agents.json, agent-permissions.json, UCP, x402, NLWeb (/ask)

C-series checks follow a discover-then-validate pattern: missing endpoints drop the check rather than failing it. Don't report a missing C-check as a problem unless the user explicitly wants that protocol.

Guidelines

• Don't restate the entire spec. When explaining a failure, name the check (e.g. "P11 — JSON-LD missing top-level BreadcrumbList"), summarise what it measures in one sentence, then jump to the fix. The full Vercel Spec and llmstxt.org spec are linked from https://agent-ready.dev/methodology.
• Markdown-mirror failures cluster. P15–P20 typically fail together (no .md mirror at all). When you see all six failing on a page, recommend creating the .md route first — the rest cascade.
• Page-level vs site-level fixes. A failure on the home page might require a one-line change; the same failure across every page might be a layout-level change. Surface the difference: "this failed on 1 page" vs "this failed on every page scanned" matters for prioritisation.
• Score deltas matter more than absolute scores. If the user is iterating (scan → fix → re-scan), a +5 point movement after a single fix is the signal — call it out.
• Don't fabricate scan ids. If the user references a scan id you don't have evidence of, call get_scan first. A 404 is a clearer error message than a hallucinated scan.

When NOT to Use This Skill

• General SEO audits (page speed, Core Web Vitals, keyword optimisation, backlink analysis) — use a Lighthouse or Ahrefs MCP wrapper. Agent Ready measures AI-readability specifically, not traditional SEO.
• Performance profiling (load time, JavaScript bundle size, Cumulative Layout Shift) — use Chrome DevTools MCP or a Lighthouse wrapper.
• Crawl-discovery (find broken links, generate a sitemap, audit redirects) — use a dedicated crawler. Agent Ready follows a small page-set against a check list, not a broad-discovery crawl.
• Applying the fixes the scanner surfaces — Agent Ready is read-only. Pair it with a code-editing MCP tool (or have the user edit by hand) to land the changes.
• Scanning intranet / private / authenticated pages — the scanner blocks private and reserved IP ranges at the network layer. It only works against publicly-reachable URLs.