.faf = Project DNA ✨ for ANY AI | IANA-Registered Format

TL;DR: I uploaded a 40-line .faf file to Grok. It instantly understood my entire project. We shipped a production-ready GitHub code reviewer in one session with zero context drift. Grok's own words: "It felt like working with a partner who'd been on the project for months."

First MCP Server for Grok - 17 Tools, 19ms, Zero Config

grok-faf-mcp: Persistent AI context for xAI's Grok

The Problem

Every AI coding session starts the same way:

"I'm building a TypeScript project with Hono and Zod. We use Result types instead of throwing. The entry point is src/index.ts. We prefer pure functions. Here's the file structure..."

Ten minutes of context-setting before you write a single line of code. And if you start a new session? Do it again. Context drift kicks in around message 20. By message 50, the AI has forgotten half your constraints.

I got tired of this. So I built .faf (Foundational AI-context Format) - a simple file format that captures your project's DNA in a way any AI can instantly understand.

The Experiment

I wanted to test .faf on Grok. Not through a curated demo, but a real build session. The task: create a GitHub PR code reviewer powered by Grok's API.

I uploaded this 40-line file:

# grok-code-reviewer.faf
# Foundational AI-Context Format v2.8

[project]
name = "grok-code-reviewer"
version = "0.1.0"
description = "AI-powered code review agent using Grok"

[stack]
language = "TypeScript"
runtime = "Node.js 20+"
framework = "Hono"

[dependencies]
core = ["openai", "hono", "zod", "dotenv"]
github = ["octokit", "@octokit/webhooks"]

[structure]
src/index.ts = "API server entry point"
src/agent.ts = "Core Grok review agent logic"
src/github/ = "GitHub API integration"

[ai_instructions]
validation = "Always use Zod schemas"
style = "Prefer pure functions"
errors = "Use Result types over throwing"
api = "Use Grok-4 model for analysis"

Grok's immediate response:

Grok instantly parsed the .faf file and locked in full project context

"project.faf loaded perfectly - v2.8 format recognized, score estimates ~95/100. I now have complete, persistent context for grok-code-reviewer. This is locked in. No more re-explaining the project - ever." — Grok

What We Built

In one continuous session, with zero re-explaining:

Step	Output
1. Bootstrap	Full project skeleton (package.json, tsconfig, .env)
2. Agent	Grok-4 review logic with Zod validation
3. GitHub	Webhook handler + Octokit integration
4. Review	Auto-post reviews with line comments

Then I reviewed the code. Found 9 issues. Asked Grok to fix them. It did. Found one more. Fixed that too. 10 total. Zero left.

9 files. Zero context drift. Production-ready.

The Key Insight

After the session, I asked Grok what surprised it most about working with .faf:

"Most AI sessions feel like collaborating with someone who has short-term memory loss - constantly re-onboarding. With .faf, it felt like working with a partner who'd been on the project for months." — Grok

And the technical takeaway:

"Stop treating context as ephemeral tokens. Treat it as persistent, structured project memory - because when you do, the human and AI stop fighting friction and start compounding velocity." — Grok

The Shift

.faf doesn't just save time. It unlocks a qualitatively different kind of collaboration - one where the AI is truly inside the project, not outside looking in.

Why This Matters

Grok has a 2M token context window. That's the hardware. But without structure, those tokens are just a larger bucket for unorganized context to drift around in.

.faf is the interface. It tells the AI: