Operating Stack · how I direct AI
The layer that makes one operator run like a team.
The model is the engine. This is the operating system I run it with.
Everything else on this site — the repos, tools, cron jobs, architectures — is the output. This section is the method: a versioned operating doctrine, an execution register for the work that has to land, the hooks that keep both live automatically, and a live instrument that shows me exactly what every session costs and where it stands.
The four pieces
Operating doctrine
20 behavioral clauses I run the model against — versioned (NEXUS v37.0), injected into every session.
OpenThe Claude HUD
A VS Code extension I built — five live gauge cards for the session in the window: context pressure, account limits, cache warmth, and a per-subagent breakdown.
OpenThe status line
Five figures, and only the 5 the HUD cannot see. Zero subprocesses per render.
OpenThe hooks
32 hooks across 6 Claude Code events — the layer that makes a convention something the tool enforces rather than something I remember.
OpenThe Fable register
RETIRED — a frontier-execution mode for work where the finished quality was the whole deliverable.
Read the archiveWhy bother
A frontier model is only as good as the operator directing it. The gap between a chat and a shipped system is discipline — spec first, verify adversarially, never claim a result you didn’t observe. I encode that discipline once, in NEXUS v37.0 and a set of hooks, so it runs on every task without being re-typed. This is that layer.
Data reviewed · current · kept by hand