Faff — AI skill for Claude Code

SuperDomestique (formerly known as `faff`)

SuperDomestique enables progressively autonymous AI software delivery by combining AI-led execution with deterministic governance, explicit engineering constraints and verifiable evidence.

It works from a tracker, moves work through specification, implementation, review, testing, and delivery, and returns unresolved decisions to a person. The aim is simple: make an agentic workload safe to stop watching, one step at a time.

`SuperDomestique` is the project currently implemented and distributed as `faff`. The `faff` repository, skills, plugin and CLI commands retain their existing names while the project evolves.

Two parts, one delivery loop

**The delivery system** turns intent into shipped work. It helps order a backlog, prepare specifications, build changes, review them, and deliver the result. The tracker remains the place where people set direction and see what happened.

**Commissaire** is the governance system. It checks the evidence produced by a run, applies deterministic rules at named gates, and stops or parks work when the available evidence is not enough to continue.

Both are part of the current `faff` repository and distribution. Commissaire is a logical code boundary today, not a separate service or security boundary.

How far can you step away?

Autonomy is assigned per workload. A team can keep sensitive work at L1 while letting routine, well-covered work run at L3 on the same board.

Level	Working relationship	Current position
L1	You do the engineering; SuperDomestique helps plan and specify it.	Available
L2	An agent builds one change and waits for approval at the major gates.	Available
L3	An agent drains eligible work unattended; ambiguity is parked for a person.	Current unattended path
L4	A run is intended to proceed without scheduled supervision.	Preview; mechanisms exist, external proof is incomplete

The [levels guide]