Overview

The Atlas

Making tacit knowledge shared — between humans and AI agents.

The agent writes the code. You hold the understanding.

Every experienced developer carries an understanding of their system that goes beyond documentation and code. The components, how they connect, where the boundaries are — knowledge built up over time, stored nowhere but grey matter. When an AI agent writes most of the code, that understanding becomes the most valuable thing the human has. And the hardest thing to share.


The Problem

Tacit knowledge is trapped

Every experienced developer carries tacit knowledge of their system — the components, how they connect, where the boundaries are. Not documentation. Not code. The understanding itself (noesis). It's what lets you hear a proposed change and instantly know whether it fits or breaks something.

Ctrl/Cmd + wheel to zoom · Scroll to pan · Drag when zoomed · Double-click to fit

Loading...

That knowledge is invisible, trapped in grey matter, impossible to share:

Yours is isolated

Your mental model might be brilliant — ahead of the code, mid-refactor in your head — but teammates can't see it. The agent can't see it. Multiple people on the same project carry different versions and don't know it.

The agent has none

It re-derives structure from code each session. Full architecture docs don't fit in context permanently. Without a compressed, loadable representation, every session starts from scratch.

Newcomers start cold

New humans, new agent instances, you after a month — no way to absorb the grey matter directly. Onboarding is archaeology.


The Thesis

Externalise it

Make the tacit knowledge concrete, shared, and loadable. When both human and agent operate from the same compressed understanding, the collaboration changes fundamentally:

The human steers from a shared reference

Not from memory alone. Reviewable, challengeable, versionable. When you disagree with the agent's direction, you can point at the same picture and say here.

The agent reasons with precision

It stops re-deriving and starts reasoning like someone who's been on the project for years. Every session. Context architecture — the understanding injected so reasoning is grounded, not reconstructed.

Alignment becomes visible

When your picture and the agent's picture diverge, you can see it. Misalignment isn't assumed away — it's surfaced and resolved.

That's not a documentation problem. It's a collaboration problem — and the solution is an artifact that both sides read, write, and trust.

The Artifact

A picture that speaks a thousand lines of code

A compressed, visual representation of the system — diagrams, orienting prose, source-linked references. Not an exhaustive architecture document. Not a wiki. A representation optimised for loading tacit knowledge quickly and reasoning against it precisely.

Both the human and the agent read and write the same pages. No "human version" and "agent version."

For the agent, it's context architecture — the understanding injected into every session so reasoning is grounded, not reconstructed. For the human, it's the same thing it's always been: the picture in your head, except now it's outside your head and someone else can see it.

For the humanFor the agent
The mental model, externalisedContext architecture — loadable understanding
Review changes against the shape, not the codeReason with the precision of long tenure, every session
Challenge, version, share with teammatesGround planning and proposals in shared knowledge

Staying Honest

An artifact that lies is worse than no artifact

The obvious failure mode: the code evolves, the representation doesn't, and now you're steering against stale knowledge. Trust erodes the moment the artifact falls behind.

The solution is automated drift detection:

1. Structural scan
Git diff since last check — new directories, infrastructure changes, entry points. Cheap, no LLM.
2. Agent evaluation
If structural changes found, an agent reads the artifact and the diff. Decides what's stale and why.
3. Update PR
Agent updates the representation, opens a PR. Human reviews and merges — or rejects.

The artifact stays honest automatically. The human stays in the approval loop via PRs, not manual auditing.

An honest snapshot is worth more than a precise but stale one. The goal isn't real-time accuracy — it's trustworthiness as a shared mental model.

Who It's For

Grey matter, shared

You and your agent

On any codebase. Brownfield — the agent reverse-engineers it from existing code, proposes it, you confirm or correct. Greenfield — it grows alongside the system. Returning after a month — 10 minutes to reload the grey matter.

Your team

The tacit knowledge that lived in one person's head is now visible, shared, and versionable. New people don't do archaeology — they absorb the compressed understanding directly. How fast they swim depends on their experience, not their tenure.

Anyone building with AI agents

Most AI coding tools give you a faster typist. This gives you a collaborator that shares your understanding of the system — and keeps it honest when the code changes.


The Philosophy

The agent sees the trees. You see the forest.

The atlas is the map in between — the tacit knowledge made concrete, the grey matter made shared. Without it, you're steering from memory and the agent is building without context.

Not a faster typist. A collaborator that maintains a shared understanding of your system — and keeps it honest when the code changes.

That's not a feature. It's a philosophy of collaboration.