Agent Ready

Turn a public agent-readiness score, or a log of agents failing to navigate, into shipped, verified HTTP and docs changes.

• IS: ingesting Mintlify Agent Score / AFDocs, Is Agentic, Is It Agent Ready, or url-discovery-bench findings, plus server-side evidence of agent 404s, inspecting the repo, implementing the matching protocols, adding tests, and verifying live responses.
• IS NOT: rewriting docs prose (ghostwriter), package/CLI/SDK ergonomics (dx-audit), whether an in-product agent can be trusted (ax-audit), search ranking, crawler policy, and Next.js llms.txt/markdown routes (seo; this skill owns the AFDocs and Is Agentic contract those routes must satisfy), or repo instruction files (agents-md). Do not vendor vercel-labs/is-agentic; that skill retrieves reports. This one implements the product.

Contents

• Workflow
• Reference files
• Priority
• Output
• Gotchas

Workflow

Agent-ready progress:
- [ ] Step 1: Ingest the report, the agent 404 log, or run scanners against the public URL
- [ ] Step 2: Inspect the existing codebase before any edit
- [ ] Step 3: Inventory every knowledge surface the origin serves and map each finding to a real one; skip surfaces the product does not offer
- [ ] Step 4: Load the matching reference and implement the map first, then failures, then warnings. For Next.js App Router `llms.txt` and markdown routes, load `seo` (that skill's `nextjs-implementation.md` and `answer-engines.md`) instead of a second recipe.
- [ ] Step 5: Add or update tests for every behavior you change
- [ ] Step 6: Verify every public endpoint and machine-readable file you touched (`scripts/check-surfaces.sh`, then the curl recipes for anything it does not cover)
- [ ] Step 7: Report changes, quoted verification, and remaining product decisions

A pasted scorecard is the spec. If none is present and the user named a public URL, gather one:

pnpm dlx afdocs check <docs-url> --format scorecard --sampling deterministic
pnpm dlx is-agentic <domain> --json

Is It Agent Ready: POST https://isitagentready.com/api/scan with {"url":"<origin>","format":"agent"}. Prefer the user's pasted report over a new scan when both exist.

Server logs are a brief too. Agents fetch server-side and run no JavaScript, so they never appear in client-side analytics; count requests for .md URLs, llms.txt, and recognized AI user agents, and treat a stream of 404s from those agents (fetch, 404, guess a sibling path, retry) as the failing check. references/verification.md has the log queries and the navigation benchmark.

Step 3 covers the whole origin, not only /docs: changelog, release notes, help center, community, and status pages are knowledge agents answer from, and they are usually HTML-only while the docs are ready. Decide per surface whether it gets markdown twins and an index entry; marketing pages do not.

Preserve visual design and existing product behavior. Change discovery, representations, headers, and documented contracts, not the feature set.

Local test suites that cannot reach production are safe to run, fix, and rerun. Do not deploy, change DNS, buy a registry name, or write outside the working tree without authorization.

Done when every in-scope failing check has a code or content change (or an explicit skip with reason), tests cover the new behavior, and Step 6 quotes status, Content-Type, and the relevant headers or body from the environment you actually hit.

Reference files

Priority

Agents read markdown fine and fail at navigation. In Mintlify's 2026 benchmark (2,400 tasks, 20 docs sites, Claude and Codex) accuracy held at 94 to 99% in every format while failed requests per task went 2.23 on HTML, 1.42 on plain markdown, 0.11 once each markdown page linked llms.txt. Order work accordingly:

1. The map: llms.txt on the docs host, a link to it in the first lines of every markdown twin, and Link headers that advertise both. Agents request .md and llms.txt only when they know they exist.
2. Failures the product actually has (docs HTML that agents cannot read, HTML error pages on a real API, gated public docs with no alternate path).
3. Warnings on those same surfaces (llms.txt coverage, buried directives, wrong Content-Type).
4. Recommended checks that match a surface already in the repo (OpenAPI, MCP, OAuth).
5. Emerging extras (commerce protocols, A2A, DNS-AID) only when the product already offers them or the user asked to add them.

A missing MCP card is not a failure on a site that has no MCP server. Do not invent an API, CLI, or payment protocol to chase points.

Output

Group work by check. For each: evidence from the report, files changed, exact markup or schema added, verification quote, skip reason if N/A. When the brief was a 404 log or benchmark, quote failed requests per task before and after.

Finish with remaining items that need a product decision, DNS access, or credentials.

Gotchas

• A .md twin has no navigation: the conversion stripped it with the chrome. Emit the llms.txt link from the twin route so no page can miss it; a per-page edit drifts.
• Markdown twins are a route in one app. When the changelog or blog lives in a separate marketing app on the same origin, it stays HTML-only without anyone deciding that; walk every app, not only the one with the twin route.
• One llms.txt advertised three ways (rel="llms-txt" from a proxy, rel="https://llmstxt.org/rel/llms-txt" from framework headers, rel="describedby" from a layout <link>) leaves the scanner reading whichever layer wins on that response. Pick the rel set the scanners probe and emit it from one place.
• llms.txt that lists HTML while .md twins exist steers agents away from markdown and is scored worse than linking .md from the start.
• A directive in <head>, nav, or past 50% of the HTML body does not count. Put it in the document body, near the top, server-rendered.
• A dashboard with no agent traffic proves nothing: agents run no JavaScript, and Search Console and SERP tools count searchers, not agents. Server logs of .md, llms.txt, and AI user agent requests are the only readership measure; without a log drain the honest number is No data, not a client-side proxy.
• Bot protection tuned for crawlers (challenge pages, WAF bot rules, tight rate limits on text/markdown routes) blocks the agents you are optimizing for, and they cannot pass a challenge. Exempt the machine-readable routes or serve them from paths the rules do not cover.
• Coverage fail vs curated index: regenerate from the sitemap when the site intends parity; if the index is intentional, say so and do not pad it with marketing URLs.

Maintenance only: evals/evals.json is for changing this skill, not for a user task.