AXO / Agent Readiness Layer Skill

Purpose

Use this skill to transform a normal human-facing website, business brief, product, API, or codebase into an Agent Readiness Layer: a structured, machine-readable, tool-ready, safety-aware layer that allows AI agents to discover, understand, compare, recommend, and operate the business accurately.

This skill is not only SEO, GEO, AEO, README generation, or llms.txt generation.

It is AXO: Agent Experience Optimization.

AXO answers:

• Can an AI agent discover this business and classify it correctly?
• Can an AI agent find the machine-readable layer without being told that /llms.txt exists?
• Can an AI agent understand the business without relying on visual UI?
• Can an AI agent identify confirmed facts, missing facts, offers, services, prices, locations, and CTAs?
• Can an AI agent compare this business against alternatives without hallucinating?
• Can an AI agent choose the correct next action for different user intents?
• Can an AI agent use forms, APIs, tools, browser workflows, or MCP servers safely?
• Can an AI agent know when to stop and ask for human approval?
• Can the business audit and evaluate agent behavior afterward?

When to Use This Skill

Use this skill when the user asks to:

• Make a website LLM-friendly, AI-agent-friendly, machine-readable, or agent-operable.
• Generate llms.txt, llms-full.txt, markdown mirrors, or AI-readable business documentation.
• Create README/documentation for an AI agent, Codex, browser agent, voice agent, sales agent, support agent, search agent, or internal operations agent.
• Optimize a website for AI agents while preserving normal human-facing pages.
• Build structured business docs, schema, metadata, conversion rules, API/tool readiness maps, permission models, or agent evals.
• Audit whether a website can be discovered, understood, recommended, or operated by agents.
• Prepare a business for MCP, OpenAPI/function-calling, browser agents, voice agents, or autonomous workflows.

Do not use this skill for simple SEO metadata only unless the user specifically wants the agent-readiness layer included.

Core Philosophy

A human website sells. A machine-readable layer explains. A discovery layer helps agents find and compare. A discovery hardening layer makes the machine-readable layer findable from a cold crawl. A tool layer lets agents act. A permission layer keeps actions safe. An evaluation layer proves the system works.

Use the eight-layer model:

1. Eyes — Can the agent understand?
2. Discovery — Can the agent find, classify, compare, and recommend?
3. Context — Can the agent retrieve the right source of truth efficiently?
4. Hands — Can the agent act?
5. Permits — Can the agent act safely?
6. Brain — Can the agent decide correctly?
7. Memory — Can the agent stay consistent over time?
8. Evaluation — Can we prove it works?

Required Operating Rules

1. Never hide a whole website inside alt text, meta descriptions, invisible text, or spammy markup.
2. Keep the human-facing website and machine-readable layer consistent.
3. Separate confirmed facts from marketing claims.
4. Mark unknown facts as Not provided.
5. Never invent prices, guarantees, awards, certifications, reviews, opening hours, medical/legal claims, or availability.
6. If the business is medical, legal, financial, child-related, housing-related, employment-related, or otherwise high-stakes, add extra caution, escalation, and claim-control rules.
7. Use Markdown, JSON, JSON-LD, OpenAPI-style schemas, MCP-style descriptions, and plain text wherever possible.
8. Every CTA must be explicit and traceable to a confirmed source.
9. Every high-risk action must have a permission level and human approval rule.
10. Every generated output should make it easier for an agent to answer: who is this for, what do they offer, where are they, how do I contact them, what should I recommend next, what can I do, what should I not claim, and when should I escalate?
11. Design for both API agents and browser/computer-use agents.
12. Do not treat llms.txt, llms-full.txt, or /docs as complete until they are discoverable from normal crawl paths.
13. A machine-readable layer is incomplete until it is discoverable from at least three independent paths: homepage HTML, sitemap/robots, and a visible or crawlable documentation index.
14. Every advertised machine-readable route must return 200 or redirect intentionally. Never link agents to a 404.
15. Generate evals before treating the system as ready.

Inputs to Request or Extract

When available, extract these from the website, brief, codebase, provided copy, sitemap, screenshots, API docs, product docs, or user notes:

• Business name
• Industry
• Location and service area
• Contact methods
• Opening hours
• Languages
• Primary CTA
• Secondary CTA
• Products/services
• Prices/packages
• Booking or buying process
• Target audience
• Pain points
• Differentiators
• Trust signals
• Reviews/testimonials, only if real/provided
• Legal/medical/financial constraints
• Existing pages and URLs
• Existing forms and fields
• Existing APIs or integrations
• Existing schema and metadata
• Existing internal docs or README
• Existing agent/tooling architecture, if any
• Existing CRM, booking, payment, email, analytics, or support systems
• Existing auth/security constraints
• Existing observability, logs, and QA/eval systems

If information is missing, do not ask endless questions by default. Create a Missing Information Report and continue with best-effort placeholders labeled Not provided.

Workflow

Phase 1 — Source-of-Truth Extraction

Create a factual business profile before generating any docs.

Output:

• Business Source of Truth
• Fact Confidence Map
• Missing Information Report
• Risk Classification
• Claim Control List

Use templates/business-profile.md, templates/source-of-truth.md, and templates/missing-info-report.md.

Phase 2 — Human Website Structure Audit

Check whether the human website is clear, conversion-focused, accessible, and structurally understandable.

Assess:

• Semantic HTML structure
• Heading hierarchy
• Page purpose clarity
• CTA clarity
• Form labels and input names
• Contact visibility
• Mobile navigation clarity
• Service/category structure
• Trust and proof placement
• Accessibility basics
• Whether critical business facts are visible to humans, not hidden only in machine docs

Output:

• Human UX gaps
• Conversion gaps
• Structural gaps
• Accessibility gaps

Use checklists/human-website-audit.md.

Phase 3 — Eyes Layer: Machine Readability

Create or plan the machine-readable layer.

Assess or generate:

• llms.txt
• llms-full.txt
• /docs/ and /docs/index.md
• /sitemap.md
• Markdown mirrors for important HTML pages
• README.md
• /docs source-of-truth files
• Schema.org JSON-LD
• Homepage alternate markdown links
• Footer or documentation links to the machine-readable layer
• Metadata map
• Alt text map
• Sitemap and canonical URL logic
• Structured page summaries
• Machine-readable contact, offer, service, and location facts

Output:

• Machine Readability Score
• Missing machine-readable assets
• Proposed files and snippets

Use templates/llms.txt.md, templates/llms-full.txt.md, templates/readme.md, templates/page-machine-profile.md, snippets/json-ld-patterns.json, and checklists/machine-readability-audit.md.

Phase 4 — Discovery Layer: Agent Discoverability and Recommendation Readiness

Map how an AI agent discovers, evaluates, and decides whether to recommend the business.

This phase has two distinct parts:

1. Recommendation readiness — can the agent classify, compare, and recommend the business accurately?
2. Discovery hardening — can a cold agent find the machine-readable layer without being told where it is?

Assess or generate:

• Agent discovery profile
• Agent journey map
• Category and entity classification