DEV Community

Wan Satya
Wan Satya

Posted on

Are We Recreating the Pre-Git Era for AI Coding?

Rune Context

Every AI coding tool seems to invent its own memory format.

Today we have:

  • README.md
  • AGENTS.md
  • CLAUDE.md
  • Cursor rules
  • Copilot instructions

All of them try to solve the same problem:

How do we preserve understanding of a codebase?

Meanwhile, every coding agent repeatedly rereads thousands of lines of code and rebuilds the same understanding from scratch.

That feels strangely familiar.

Before Git became the standard, version control was fragmented. Every team had its own way of exchanging patches and managing history.

Git changed that.

.git/ → what changed
Enter fullscreen mode Exit fullscreen mode

But repository understanding is still fragmented.

Source Code Isn't Understanding

Humans rarely understand a codebase by reading every file.

We rely on:

  • architecture
  • dependencies
  • conventions
  • feature boundaries
  • tribal knowledge

AI agents should probably do the same.

What If .git/ Had a Sibling?

I've been exploring the idea of a .rune/ directory:

repo/
├── .git/
├── .rune/
├── src/
└── tests/
Enter fullscreen mode Exit fullscreen mode

Where .rune/ contains things like:

.rune/
├── spec.md
├── architecture.md
├── graph.json
├── files/
└── features/
Enter fullscreen mode Exit fullscreen mode

Not source code.

Not embeddings.

Not hidden state.

Just repository understanding stored as plain files.

.git/   → history
.rune/  → understanding
Enter fullscreen mode Exit fullscreen mode

Why Not AGENTS.md?

AGENTS.md works.

But every tool is creating its own format:

  • AGENTS.md
  • CLAUDE.md
  • Cursor rules
  • Copilot instructions

These are useful, but they're mostly tool-specific.

I'm wondering whether repository understanding itself should become portable and versioned, in the same way Git standardized source control.

Not Another Agent Framework

Rune isn't trying to replace:

  • MCP
  • vector databases
  • coding agents

Those solve different problems.

The goal is much simpler:

Make repository understanding explicit and shareable.

Maybe I'm Wrong

Maybe AGENTS.md is already enough.

Maybe embeddings solve this problem.

Maybe repository understanding doesn't need its own layer.

But it feels like we're still in the "pre-Git era" of AI coding, where everyone is solving the same problem differently.

Curious what others think.

Do we need a standard for repository understanding, or am I overthinking this?

Git stores history.

What should store understanding?

https://github.com/rune-context/rune/wiki

Top comments (0)