in production since 2026-06-03
One designer.
Nine agents.
A working practice.
os.Claude is a six-layer operating model that runs a real design practice on Claude Code. Agents triage, build, review, and merge. Routines run overnight. Every skill, convention, and profile that makes it work is documented here — and free to take.
16
agents and operators
154
canon artefacts
388
pull requests merged
128
days in production
What this is
Not a product. A practice, made legible.
It runs. This is not a concept or a demo. The agents on this site ship the software, sweep the pipeline, plan the week, and write the morning brief for a working design practice — every day, on schedules.
It is documented. Every behaviour is written down as an artefact: skills, tasks, routines, agent profiles, conventions. The system reads its own documentation to run; the site is generated from the same source.
It is yours to take. All 154 artefacts are downloadable as plain markdown. Use them as a starting point for your own operating model — the ideas hold beyond one practice.
Layer 03
The fleet
Each agent is a named role with a mission, a remit, and hard boundaries. No human names — an addition to the ways of working, not a person replaced.
Layer 04
How a ticket becomes shipped software
01
Relay triages the ticket and routes it
02
Forge builds it and opens a pull request
03
Prism reviews it and writes the verdict
04
Relay merges on approval, marks it Done
The operator gates what matters; a clean pass routes itself. QA bounces are capped at two strikes before a human looks. Overnight routines drain the backlog on cheap tokens.
The model
Six layers, two of them read-only
The story
"The unlock wasn't automation. It was writing the practice down precisely enough that something else could run it."
Open canon
Take the artefacts
Every skill, agent profile, convention, and routine — plain markdown, no sign-up, no licence theatre. If it helps you build your own operating model, it has done its job twice.