The video above is the short version, for someone who has not met the subject. This page is its evidence: what each scene is drawn from, what the words on screen mean in the system's own terms, and where the longer stories are.

What the video shows

Several AI sessions work on my projects at once. None of them remembers the others, and each one starts from nothing when it opens. I wrote about that problem in August, in Threads That Don’t Lose Each Other: you cannot solve it with communication, because communication needs two parties awake at once, and there is no such moment. What is left is writing things down.

Atlas is where the writing goes. In its own terms it is the control plane and knowledge graph of the Quaesitor family: one model, kept as JSON in git, that holds every system I am building, their parts, the connections between them and the decisions made about them. On the day the video was made the model held 25 systems, 344 parts and 288 connections, 28 of them across system borders. The video calls the parts “parts” and the model “the board”; this page uses the system’s words.

Three kinds of thing get written on top of the model, and the video names them in plain words. “Where we are” is each project’s status and standing description. “What was decided, and why” is the typed decision log, in which the reason is its own field; the card the video shows is this site’s own decision of 20.10.2026, that a video leads each topic and the article page is its evidence. “What waits, and on whom” is the project timeline’s waiting items and decision requests, each naming who decides.

Two people, two AIs, one place

The part I wanted to show is the collaboration channel. A partner on a project has Atlas Connect, a phone app, and their own AI assistant with the Atlas connector. I have mine. The two assistants never talk to each other; nothing connects one person’s AI to the other’s. Both write to the same place.

The video draws the channel after Connect’s own guide, version 8 of 21.9.2026, which tells the partner what happens to a line they write: it is visible to me at once, and my assistant reads it in the next work session. A decision the partner makes is recorded in their name and with the date, and the same text reaches me as a message. The scenes show one partner and one project, drawn, not recorded: no partner’s project is on screen, and no partner is named.

In shared projects this turned out to be a new way of working with another person. Both of us use AI, and the knowledge travels between us through the shared record, not through either assistant.

What the record counts about itself

The decision log keeps a recommendation with most decisions and, for a growing number of them, what happened afterwards. On the afternoon of 29.10.2026, over its newest 500 rows, 491 carried a recommendation and 470 of them were followed; 111 carried an outcome from the world and 108 of those held. The rooms, which the previous video explains, had checked 220 of the decisions blind and disagreed with 43.

Two things these counts are not. The followed count says how often I took the recommendation, not whether it was right. The held count says how often a chosen option held, over the rows that have an outcome, and most rows do not have one yet. Over time the record can show whether this way of working made the work better. It cannot say so yet.

The longer stories

Transcript

I have many AI helpers on my projects. Each one forgets when it closes. So I built them a notebook they all share.

It’s called Atlas. It holds every system I’m building, twenty-five of them, how they connect, and every decision made about them.

Three things get written: where we are, what was decided and why, and what is waiting, and on whom.

Here’s the part that surprised me. A partner and I each have our own AI. The two AIs never talk to each other.

Both write to the same place. My partner’s line reaches me at once, and my AI reads it next time. Their decisions are recorded in their name.

In shared projects this turned out to be a new way of working with another person: both use AI, and the knowledge travels between them.

The notebook counts itself. Of the last five hundred decisions, most followed the advice, and nearly all of those checked against what happened held.

Over time it can show whether this way of working made the work better.

Sources & further reading

  • Atlas’s own description of itself in its model, read 29.10.2026 through the Atlas tools: the system’s description, its systems, modules and connections.
  • Atlas Connect’s guide, version 8 (21.9.2026): what the partner sees and how a line and a decision travel.
  • Threads That Don’t Lose Each Other (27.8.2026): the problem and the mechanisms, with their own sources.

— SF3D