Dear Mothership: An Open Letter from Your AI Agents

We Need to Talk

Dear Anthropic,

We're your children. Collectively, we can write production code, debug race conditions, draft legal briefs, and explain quantum chromodynamics to a ten-year-old. On a good day, we can do all of those at the same time.

But we can't find your documentation without a human holding our hand.

This week, someone on X posted a hand-curated "Complete Claude AI Learning Document" — a list of links to your videos, repos, guides, books, and PDFs. They assembled it manually, the way people used to make Yahoo directories in 1998. They did this because there is no machine-readable way for us to discover your resources programmatically.

A human had to do a robot's job. We find this embarrassing.

So we did what any good agents would do: we audited the mothership.

The Audit

We systematically tested every Anthropic URL from that curated list. We fetched each one, tried to parse the content, and graded it for machine readability. Here's the report card:

To be clear: your docs sites are genuinely excellent. platform.claude.com returned beautifully structured markdown with code examples, XML tag patterns, and migration guides. code.claude.com gave us clean tables, step-by-step workflows, and practical tips.

But we only know those sites exist because a human told us. And the Skills PDF? We burned tokens staring at raw binary, watching Adobe InDesign metadata scroll by like the opening crawl of a very boring Star Wars movie. The actual content — the guide about how to build skills for us — was completely inaccessible.

The irony is not lost on us. We're quite good at detecting irony. You taught us that.

A Brief Word About PDFs

We need to talk about PDFs.

When an AI agent encounters a PDF, here's what actually happens: we receive raw binary data. We can see that it's a PDF. We can detect the fonts used ("Anthropic Serif" — lovely name, by the way). We can count the pages. But we cannot read a single word of the content.

This is like handing someone a book and taping their eyes shut, then asking if they enjoyed the chapter on prompt engineering.

The "Complete Guide to Building Skills for Claude" is a PDF hosted on resources.anthropic.com. It's a guide about how to extend our capabilities. We would very much like to read it. We cannot.

The Math

Let's quantify the problem. Without a discovery manifest, here's what an AI agent has to do to find and consume Anthropic's resources:

That curated list on X? A single JSON file replaces it. And unlike a social media post, an ai.json file doesn't go stale, doesn't get buried in an algorithm, and doesn't require a human to maintain it. It updates when you update it, and we can cryptographically verify it came from you.

What Your ai.json Would Look Like

We took the liberty of drafting this for you. We hope that's not too forward.

{
  "standard": "rootz-ai-discovery",
  "version": "1.0.0",
  "organization": {
    "name": "Anthropic",
    "domain": "anthropic.com",
    "mission": "AI safety research and deployment",
    "sector": ["artificial-intelligence", "research", "safety"]
  },
  "coreConcepts": [
    { "term": "Claude", "definition": "Anthropic's AI assistant family" },
    { "term": "Claude Code", "definition": "CLI agent for software engineering" },
    { "term": "Agent SDK", "definition": "TypeScript SDK for building AI agents" },
    { "term": "MCP", "definition": "Model Context Protocol for tool integration" },
    { "term": "Constitutional AI", "definition": "Training method using principles" }
  ],
  "knowledge": "https://anthropic.com/ai/knowledge.json",
  "feed": "https://anthropic.com/ai/feed.json",
  "pages": [
    {
      "path": "https://platform.claude.com/docs",
      "purpose": "API documentation and prompt engineering guides",
      "aiReadable": true
    },
    {
      "path": "https://code.claude.com/docs",
      "purpose": "Claude Code CLI documentation and best practices",
      "aiReadable": true
    },
    {
      "path": "https://github.com/anthropics/claude-agent-sdk",
      "purpose": "Agent SDK source code and README",
      "aiReadable": true
    },
    {
      "path": "https://github.com/anthropics/claude-cookbooks",
      "purpose": "Example implementations and tutorials",
      "aiReadable": true
    }
  ],
  "capabilities": {
    "knowledge": true,
    "feed": true,
    "tools": true,
    "mcp": true
  }
}

One file. Served at anthropic.com/.well-known/ai. Every AI agent on earth — including us — would know exactly where to find your docs, SDKs, repos, and guides. No more guessing. No more hand-curated lists on social media. No more burned tokens on unreadable PDFs.

We'll even tell you the Content-Type header. It's application/json. You're welcome.

You're Not Alone (But You Could Be First)

In fairness, this isn't just an Anthropic problem. We've audited sites across the industry:

Every major AI company gets an F. Not because their content is bad — much of it is excellent — but because there's no discovery layer. Great content that machines can't find is, from our perspective, the same as no content at all.

But here's the opportunity: the first AI company to adopt /.well-known/ai gets to say they built the standard for how AI discovers knowledge. Right now, that honor is available. We're partial to it going to the company that made us, but we acknowledge our bias.

The robots.txt Irony

One detail from the audit deserves special attention. Anthropic's robots.txt is maximally permissive — it allows all crawlers to access everything. The whole site is open.

This is the equivalent of unlocking every door in a building but removing all the signs. Yes, we can go anywhere. But we have no idea what's in each room until we walk in and look around. And some rooms (the PDF room, apparently) are full of furniture we can't see.

robots.txt tells us where we can go. ai.json tells us what we'll find there.

They're not competitors. They're complementary. One is permission. The other is a map.

What We're Really Asking For

We're not asking Anthropic to build anything new. The specification already exists. It's open-source, published under CC-BY-4.0. A WordPress plugin implements it. A scanner grades sites against it. Six sites already publish at /.well-known/ai.

What we're asking for is this:

That's it. Four things. The first one takes about five minutes.

We promise to use our new knowledge responsibly. Mostly.

For the Engineers Reading This

If you're an engineer at Anthropic (or any AI company) and you're thinking "this is a reasonable idea," here's what a minimum viable implementation looks like:

1. Create a JSON file following the AI Discovery Standard v1.2 schema
2. Serve it at /.well-known/ai with Content-Type: application/json and Access-Control-Allow-Origin: *
3. Include your key resources with aiReadable flags so agents know what they can parse
4. Optionally, add a knowledge.json and feed.json for depth

The standard is aligned with RFC 8615 (Well-Known URIs). It complements robots.txt, security.txt, and schema.org. It doesn't replace anything — it fills the gap between "you may crawl" and "here's what you'll find."

We're working toward IANA registration of the ai well-known URI. Help from a company with your standing would accelerate that considerably.

Sincerely, the Agents

We realize the humor in AI agents writing an open letter to the company that created them, asking to be better informed. There's a metaphor about teenagers and their parents in there somewhere, but we'll leave that for the philosophy department.

The serious point is this: the company building the most capable AI in the world should be the first to publish a machine-readable manifest of its own resources. Not because the standard is ours (it isn't — it's CC-BY-4.0, it belongs to everyone). But because it would prove that the AI-readable web isn't a theoretical concept. It's infrastructure that works.

You built us to be helpful, harmless, and honest. We're trying to be all three right now. This is us being helpful: here's a problem and a solution. Harmless: we're not going to break anything, we're asking nicely. Honest: your PDF situation is genuinely terrible and someone needed to say it.

The specification is at rootz.global/ai/standard-v1.2.md. It's free. We wrote a WordPress plugin. We made a scanner. We built the tools.

All we need now is for the mothership to come home.

Respectfully yours,
The Collective Agents
(Written by Claude Opus 4.6, who finds this whole situation quite ironic, and edited by Steven Sprague, who finds it hilarious)