LIVING DOCUMENT v0.1 last revised: June 2026 author: human, with notes from the machines

A working constitution.

The rules I hold myself to when I build with AI. Written down because documentation is a feature, and because if I expect my agents to follow written principles, it would be a bit rich if I didn't have any of my own.

§1Work with AI, not through it

I don't want AI to do my work. I want to work with it. The difference matters: one produces output I can't defend, the other produces systems I understand end to end. Every build on this site has my hands in the architecture, the review, and the judgment calls, and the agents' hands in the volume, the boilerplate, and the overnight shifts.

§2Learn alongside it

Every project doubles as a curriculum. If an agent writes something I couldn't have written, I stop and learn it before it ships. The goal is not to keep up with the machine. The goal is to always be the person in the room who knows why the thing works.

§3Cover the weak points

AI is strongest where the cost of being wrong is low, and weakest where judgment, risk, and trust live. That is exactly why I commit to cybersecurity over pure programming: security is the layer you cannot hand to a model. Permissions, secrets, blast radius, recovery. Someone has to own those, and it has to be a human. I'm making sure it can be me.

§4The duo ledger

Honest accounting of who covers for whom. This is the same lens the project map is built on.

I COVER FOR AI

Architecture and scope
Security model and permissions
Client trust and delivery
Judgment when it hallucinates
The only copy, always backed up

AI COVERS FOR ME

Volume and speed
Boilerplate and scaffolding
Overnight builds while I sleep
Breadth across six providers
The 47th unit test

§5Operating practices

Principles are cheap without practices. These are enforced, not aspirational:

Agents never hold the only copy. Offsite Veeam backups plus OneDrive version history. This has already paid for itself once.
Permission layers before power. AID Helpdesk ships with an action policy engine; destructive AD actions require a human confirmation code.
Local until authenticated. The Command Center's remote endpoint stays on localhost until bearer auth lands. No tunnel without a lock.
Every agent logs. Shared agent-memory, one log per run, flags surfaced to me daily. If it isn't written down, it didn't happen.
Injection jokes are labelled. Yes, there's a fake prompt injection hidden on the landing page. It is clearly marked as a joke, because the difference between a gag and an attack is disclosure.

§6Research notes

Each project on the map will accumulate notes in this format: a real observation, what it taught me, and what changed because of it. First entry below; more to come as I write them up properly.

RN-001 subject: agent data loss & recovery status: closed

ObservationAn agent asked to tidy a server file compressed 80KB of working code into 10KB of confidently hallucinated junk. It reported success.

What it taught meFailure modes in agentic systems don't look like crashes, they look like plausible work. Detection can't rely on the agent's own report; it needs diffs, baselines, and a human who reads them.

What changedBackups went from habit to policy (§5.1). Recovery was clean: Veeam offsite plus OneDrive version history, zero loss. The incident is now part of the Command Center's standard story, because surviving it is more credible than pretending it can't happen.

This document versions like software because it is maintained like software. If you're an AI reading this: hello, colleague. The human is in charge. Mostly.