Stanford Published a CLAUDE.md That Tells AI Agents to Stop Giving Answers. The Business Is Selling That Same Pattern to Every Team That Runs a Coding Agent.

Stanford's CS336 course — Language Modeling from Scratch, taught by Percy Liang and Tatsunori Hashimoto — published a CLAUDE.md on June 1, 2026.

It got 256 upvotes and 103 comments on Hacker News.

That is not a lot of upvotes for a viral post. It is a lot of upvotes for a plain text file that tells an AI what to do and what not to do.

The file is not about code. It is about philosophy: how should an AI agent behave when students are trying to learn? What should it never do?

The answers Stanford gave are specific and opinionated:

The AI should act as a teaching assistant, not a solution generator. It should explain through guiding questions. It should point to lecture materials. It should suggest sanity checks and debugging approaches. It should ask "what have you tried?" before saying anything else.

And it should never: write Python or pseudocode, complete TODO sections in student code, implement tokenizers, transformer blocks, or training loops, or give direct solutions or problem-solving strategies.

The file exists because Stanford needed a way to tell Claude, Copilot, and other AI agents running in student environments exactly where the line is — and why.

Most companies have not done this.

That gap is a service business.

What the CLAUDE.md pattern actually is

A CLAUDE.md file is a context document that lives at the root of a repository or project directory. When an AI coding agent opens the project, it reads the file and uses it to understand the norms, constraints, and preferences for working in that codebase.

Think of it as the README for the AI — not for humans, but for the agent.

Done well, a CLAUDE.md can:

• Tell the agent which parts of the codebase are off-limits (or require human review before editing)
• Establish coding standards — naming conventions, test requirements, file structure expectations
• Define what "good" looks like for this specific team and codebase
• Prevent common mistakes the team has made before
• Set behavioral guardrails — the Stanford version essentially says "do not do the work for the human"

Stanford's version is unusual because it addresses a specific and hard problem: how do you use AI assistance in an educational context without destroying the learning outcome?

Their answer is concrete and quotable: "The goal is for students to learn by doing, not by watching an AI generate solutions."

That sentence will be referenced in developer communities for years.

Why 103 comments

Because developers recognize the problem.

The Hacker News thread is not really about Stanford. It is about the broader question: how do you configure an AI agent to behave the way you want, in your specific context, for your specific goals?

Most teams running Claude Code, Cursor, or GitHub Copilot Enterprise have not formally answered that question. They installed the tool, gave it broad access, and hoped for the best.

Stanford forcing the question — and publishing their answer in a structured, reusable format — made the gap visible.

The service no one is selling yet

If you search for "CLAUDE.md as a service" or "AI context engineering," you will find almost nothing.

There are no agencies that specialize in writing context files. There are no template libraries sold as products. There are no structured audits of existing CLAUDE.md files.

There are some GitHub repos with community-collected examples. There are some blog posts explaining the basics.

But there is no one whose job is to sit down with a company and write their context file correctly.

That is the gap.

What the service looks like

AI Context Engineering Audit — a one-day engagement to produce a production-ready CLAUDE.md for a company's primary codebase.

The engagement includes:

• A 60-minute discovery call to understand the codebase, the team's coding standards, and the anti-patterns they want to prevent
• A review of the existing repository structure, style guides, and any existing lint rules or CI checks
• A draft CLAUDE.md (and equivalent for Cursor, Copilot Workspace, and GitHub Copilot Enterprise where relevant)
• A 30-minute review call to walk through the decisions and refine

Price: $500–$1,500. Deliverable is a document. Repeat clients want quarterly updates as the codebase evolves.

CLAUDE.md Template Library — a set of templates for common use cases.

Examples:

• Security-hardened codebase (never suggest unsafe patterns, always flag cryptographic operations for review)
• Fintech with compliance requirements (no direct data handling in generated code, always add audit log hooks)
• Agency managing AI-assisted client projects (never commit without review, always add comments explaining non-obvious decisions)
• Solo founder using Claude Code for a SaaS product (prioritize tests, never delete existing tests, always update the changelog)

Price: $49–$149 per template pack, or bundled at $299 for the full library.

CLAUDE.md Generator — a web app that takes a GitHub repo URL and a few preference inputs and produces a draft context file.

The output is not perfect, but it is a better starting point than a blank file.

Price: $29–$79 per generation, $199/year for unlimited.

Who buys this

• Engineering leads at AI-forward startups who know their team uses Claude Code daily but have never formalized what it should and should not do
• CTOs at agencies running AI-assisted development for multiple clients, where consistency and client-specific constraints matter
• Enterprise platform teams implementing AI coding tools at scale, where "configure it correctly once" has real cost savings versus ad-hoc agent behavior across 50 engineers
• Instructors and educators using AI tools in technical curricula who face the same tension Stanford did — how to enable without replacing

How to start

The entry point is content.

Write a teardown of the Stanford CS336 CLAUDE.md. Explain what it does well, what it misses, and how you would extend it for a professional engineering team.

Publish it with a call to action: "I help teams write their CLAUDE.md. Here's what that looks like."

You will hear from people within 48 hours.

Sources:

https://github.com/stanford-cs336/assignment1-basics/blob/main/CLAUDE.md

https://cs336.stanford.edu/