Act V · Apply it
AI Factory · Playbook
Apply an AI SDLC to your own project
- 1
Write one constitution, for humans and every agent
Principles, Definition of Done, conventions, a sequential workflow, hard gates, roles, and how exceptions are granted. Make it the highest authority, and keep a small card that is always loaded.
- 2
Fork a spec-driven schema, don't invent one
Extend OpenSpec's spec-driven schema with your mandatory sections and a traceability artifact, and inject the constitution through config.yaml so every tool receives it through the same channel.
- 3
Turn every rule into an exit code
Every gate an agent must pass is a script that fails: plan completeness, live Issue check, TDD evidence, commit format, doc traceability. Run the same scripts locally (pre-push) and in CI.
- 4
Split the work into phases a model can hold
One phase per run with a fresh context, a short brief, and a deterministic gate after each phase. A failing gate feeds its own output back into a retry, up to a limit, then a human steps in.
- 5
Specialize the crew and match cost to the task
A Foreman orchestrates; specialist Workers get the narrowest skill and an explicit scope. Use cheap models for mechanical work, mid-tier for tests and routine fixes, top-tier for architecture and security.
- 6
Make the platform enforce the last mile
Branch rules that require a PR and named checks, no force-push, CODEOWNERS, and a protection ruleset kept as code and asserted by a script.
- 7
Keep humans on decisions, not keystrokes
The Owner approves specifications, grants exceptions and merges policy changes. Agents write down Owner-only steps instead of working around them.
- 8
Learn in writing
Every incident becomes a rule in a permanent document and, where possible, a test, so the next agent cannot repeat it.
What “production-ready” means here
The factory is only as good as the evidence it leaves. These are the artifacts Notaire's gates produce or check before a change counts as shipped.
Guiding principles for adapting
- Adapt, do not replace. Keep vendor engines and skills; extend them through their own extension points.
- One source of truth. The governance process lives in one document; tools point at it and never copy it.
- Agent-agnostic. Deliver the rules through a channel every tool reads (the spec CLI), not a feature of one assistant.
- Mechanical over aspirational. A rule without an exit code is a wish.