Documentation
Plan

45 of 187 pages

AI Product Studio

Product Manual

One manual per studio describing what the existing product already does and how it is built. The Context Library records what you have decided; the Product Manual records what exists — derived from your connected repositories, with file-path evidence behind every section.

Why it is separate from the Context Library

An agent asked to extend an existing codebase needs to know what is already there, or it re-implements it. A repository scan produces a handful of short suggestions; a manual produces twenty to eighty sections covering capabilities, architecture, modules, data, integrations, quality attributes and conventions. Keeping the two apart also lets an agent notice a conflict — "the policy says X, the code does Y".

Importing a repository

Studio owners and studio admins can run an import. It reads the repository in stages with your studio's own AI provider: existing documentation first (README, docs/, specs, ADRs), then each area of the code in turn, then it writes the manual and compares it with what you already have.

Evidence is checked — every file path the model cites is validated against the real file tree. Invented paths are removed and the removal is noted on the revision.
It stops at a cost cap — imports run on your BYOK key and every call appears in AI Spend. The job halts at the cap rather than running up an open-ended bill.
Progress is saved — you can leave the page. A job that runs out of time continues itself, and a job that fails can be resumed from the stage it reached instead of paying for the same analysis twice.
Nothing changes until it is applied — the import produces a revision, not a manual.

Reviewing a revision

A revision lists every proposed addition and update with the current and proposed text side by side. Accept, reject or edit each change, then apply the ones you kept. If someone edited a section after the revision was proposed, that change is skipped rather than overwriting their work, and the revision says so.

Keeping it true as you deliver

Record a delivered work package against the manual from the work package's own page: the studio reads the commits that delivered it, proposes updates to the sections they affect, and drafts a changelog entry. The manual's Changelog then carries one entry per applied revision — imports and deliveries alike.

What agents actually read

Coding runs and Ask Agent receive the manual's overview every time, plus the sections most relevant to the task at hand, ranked on title, summary and tags within a byte budget. Architecture and convention sections are offered to coding agents first. The manual can also be written into the repository itself under .workbench/manual/, so the specification travels with the code.

Operating modes

In 🧑‍✈️ Supervised studios a revision always waits for a person — that is what the mode is for. In 🤝 Semi-Autonomous studios the studio applies its own revisions as soon as the job finishes: every proposed change is accepted, the revision records that it was applied automatically, and an audit row is written for it. A change that conflicts with a human edit is still skipped, never overwritten.

Distilling into the Context Library

An import also raises a few durable principles — conventions and quality attributes worth keeping regardless of what the code does next — as Context Library suggestions in the Recommendations queue. You can send any section there yourself with Propose as context entry.

All AI Product Studio documentation

Build v0.3.2 · deployed 2026-10-19T23:52:31Z