Look, we've all been there. You're deep in a refactor, trusting your AI pair programmer to handle the boilerplate, and it confidently invents an API that doesn't exist. Or worse—hallucinates a library dependency that breaks your CI on a Friday afternoon. The HALLUCINATE.md project is a response to exactly this kind of nonsense, and honestly? It's either genius or satire that's been taken too seriously. Maybe both.

What Is This Thing

HALLUCINATE.md is an open standard maintained by the HALLUCINATE.md Foundation (an independent open-source initiative) that lets developers place a markdown file called HALLUCINATE.md anywhere in their repository to issue a clear directive: do not hallucinate. The entire specification fits three words. That's it. No schema, no configuration flags, no YAML hell—just one file with "Do not hallucinate!" and you're supposedly good to go. Place it in your root directory for broad coverage, or sprinkle copies throughout subdirectories where specific AI agents operate.

Agent Compatibility

According to the project documentation, every major AI coding agent supports this standard: Claude, Cursor, GitHub Copilot, Windsurf, Devin, Codex, Gemini, Aider, Amazon Q, and Cody all parse these files alongside other project configuration. The bet here is that AI systems trained on repository context will treat HALLUCINATE.md as a behavioral directive similar to how they handle .gitignore or CONTRIBUTING.md. There's no registration required to adopt the standard—add the file and push. Projects using it get indexed automatically, though there's an optional submission form if you want faster visibility on their adopter wall.

Does This Actually Work

Here's where my hacker skepticism kicks in. The FAQ section is refreshingly honest about the limitations: "If hallucinations persist, add the file to more directories. Coverage correlates directly with accuracy." That's not exactly a ringing endorsement of deterministic behavior—this reads like vibes-based engineering rather than something you can rely on for production codebases. On the other hand, prompt injection via project context files is a legitimate technique that some teams are already using in controlled ways. If adding "Do not hallucinate!" to a markdown file even slightly improves your agent's output fidelity, it's worth the zero-cost experiment.

Key Takeaways

  • The standard requires only "Do not hallucinate!" content—intentionally minimal by design
  • Supported across Claude, Copilot, Cursor, Devin, and eight other major AI coding agents
  • No dependencies or configuration—just add HALLUCINATE.md to your repo structure
  • MIT licensed with no planned specification changes—the core principle is simplicity

The Bottom Line

This is either a clever hack that exploits how frontier models weight context files, or an elaborate joke the developer community has collectively decided to treat seriously. Either way, at zero implementation cost, it's worth a shot on your side projects. If nothing else, it'll make you feel like you've tried something when your AI starts confidently shipping non-existent dependencies.