Arboretum
AI-Assisted
Architecting
Where direction is formed and kept
The bottleneck is no longer the code. It is the direction, and how little of it survives.
Your agents already run for hours from a written intent. The constraint moved with them, and it is no longer how fast code appears. It is how often someone can form direction worth several hours of it — and whether that direction still exists next week. The plan gets committed. The argument that produced it closes with the window.
That is what caps the height you can work at, and it is what Arboretum raises. It opens your project's own documents and lets every question, correction and decision attach to the exact paragraph, table or diagram it concerns, held as ordinary files in your own repository. When a concern turns out to span five documents rather than one, you gather those notes into a single named review, settle it once, and hand that over as one unit. One approval moves a plan, where a hundred chat messages each move a file.
What that looks like in a working project.
One project's documentation, opened in Arboretum. The workspace was prepared for these captures rather than accumulated over months; everything on screen is real.
What actually caps the height
Five answers given one at a time are five local answers.
A remark about one paragraph moves one paragraph. A settled position on how tenancy is enforced across five documents moves the system. Both cost you roughly the same minute of attention; only one of them is worth an interruption. So the height you work at is not a matter of ambition — it is set by the largest unit you are able to assemble before you have to answer.
Working question by question holds that unit at page size, for a structural reason rather than a lack of discipline. Ask about five related things in sequence and each answer is formed without sight of the other four. You get five local observations instead of one finding, and the constraint running across all five is not smaller at that width — it is absent.
The same shape makes you the bottleneck. When every local decision has to come back through the same person, everything else waits behind them, and each decision they give is formed blind to the rest. The aim is not less human judgment; judgment is the input that just became more valuable. It is to spend it on decisions that arrive in a shape worth deciding on.
That is also why durability is not a filing preference. Direction you cannot find next month has to be re-formed from scratch, usually by someone who was not there the first time — and increasingly that someone is a model with no memory of the argument at all.
What a contribution actually is
Text, plus the exact place it refers to.
A remark typed into a session is text plus whatever context the window happened to be holding. A note here is text plus a position: this block, in this section, on this page, in this zone. It also carries a purpose — suggestion, clarification, request, question, verification, decision, comment — a standing, and optionally the name of a review it belongs to.
Those are not four separate features. Together they are an address, and an address is what lets a contribution from one person and a contribution from a different assistant, made a month apart, be reconciled rather than merely piled up. Two remarks about "the auth model" collide. A note on the credential-lookup row of an enforcement table and a note on a dispatch endpoint's code block do not — they are about different things, and the record knows it.
The anchor is not a line number, which would rot on the next edit. Each note records a content fingerprint of the block it attaches to, together with its section path and the fingerprints of its neighbours, and re-resolves itself as the document changes. Where it can carry a note across an edit it does; where it cannot, it reports the anchor as ambiguous rather than pointing somewhere wrong.
The strategic part follows from that. Addressed contributions compose without a coordinator. Unaddressed ones need a person to sort them — which puts the serialization point back exactly where you were trying to remove it.
What the mechanism gives you
Four things that follow from a contribution having an address.
Wide and deep are the same dial
One block, one page, or one concern running across five documents. A question about whether a constraint holds everywhere it must can be assembled from notes on five different pages — including the inconsistency between two of them, which is visible from neither one alone.
Selection, not re-authorship
A chat can be handed a long, carefully batched prompt; nobody needs a product for that. The difference is where the batch comes from. Composing a prompt is authorship. Composing a review is selection — from situated contributions that already exist, at whatever scope the question turns out to have.
Records the project owns
Notes are ordinary files under the repository's own directory: a JSON document per page, beside the material it annotates, holding messages, purposes, standings, review names and anchors. They branch, commit and merge with the work they describe, so a review prepared on a feature branch arrives with that branch.
More than one thing to do with a scope
Hand it to an assistant. Refer to it from a terminal session and never open the application, because the records are just files. Bring in a colleague instead of a model. Record a decision so the reasoning survives being questioned in three months. Or ask for an explanation in place and choose not to keep it — persistence being a choice is part of the design, not an exception to it.
The second thing it turned out to be for
Models write documentation that is correct and hard to read.
Arboretum began as a tool for AI-assisted architecting: describing a system at the level of intent and letting implementation follow. A second purpose turned out to matter as much.
Assistants design and implement capably and then produce documentation that is technically accurate and unreadable — the structure, the emphasis, the terminology and the order of explanation each slightly wrong in ways only a person notices. Correcting that is not one edit. It is many passes, usually across a whole document set rather than a single paragraph, and it is exactly the multi-document editing loop this workspace is built for.
The same mechanism lets an assistant act as a critic rather than only an author, and lets it check its own output: with access to the rendered documentation it can look at what it just wrote instead of trusting its reading of the source. That is how a diagram definition that parses cleanly but renders unreadably gets caught.
Arboretum is the surface the rest of the work is argued on: Cheetah, Paladin, AWACS, Imprint and Troveholm each carry an installation rendering their own documents. The relationship runs both ways — it emerged from the difficulty of architecting Cheetah, then became the tool that explains and improves Cheetah, itself, and everything else. Paladin's findings are most likely to arrive on this same surface rather than in a separate product, and AWACS routes attention across the concurrent work lanes that this way of working makes possible.
Arboretum does not take the person out of the work. It lets them choose where their judgment enters it.