Sources
Rules, playbooks, truth files
Tailor
Drafts against the sources
source-traceCritic
Flags anything unsourced
Refiner
Rewrites, then loops back
Evaluation
Done means validated
voice-scanShip
Validated, every line sourced
validatedAI is an extension. Not a replacement.
My view on AI, and how I build with it. Set up right, agents extend what one person can produce. Left ungoverned, they drift and invent. So I govern them — the way you’d run a team.
Managed
Governed
Vertically integrated
25
Rules in force
4
Agent roles
0
Unsourced claims
tailorcriticrefinerevaluation
The view
Manage it like a team.
A talented hire with no brief, no review and no standards is a liability, not a force multiplier. AI is the same. The difference between value and chaos isn’t the model, it’s whether someone is directing and governing it.
So I built the governance: rules it can’t override, hooks that enforce them, playbooks that make it learn. Then I point it at real problems and let it build.
The reality
What people think AI does, and what it actually takes.
The expectation is one prompt and a finished thing. The reality is a small team of agents, each with a job, checking each other against rules they can’t override. Scroll the difference.
What people think you’re doing
Type a prompt
“Make it stunning. And don’t make it look like AI.”
Ship something great
What’s actually happening
The method
Set the agents up right.
Good output isn’t luck, it’s setup. Four parts, each doing one job: a constitution the agents obey, hooks that enforce it automatically, playbooks that make them learn, and the agents themselves, scoped and checking each other.
Brief
A posting or task arrives
Tailor
Drafts against the sources
hook source-trace
Critic
Flags anything unsourced
Refiner
Rewrites to fix
not validated ↻
Evaluation
Done means validated
hook voice-scan
Ship
Validated, every line sourced
done validated
Model
Shared by every agent
Generate, critique, refine, evaluate. The loop runs until the work validates, and nothing reaches Ship with a claim that has no source.
Rulesthe constitution
Standing policy loaded before any task: what the agents may do, must never do, and what “done” means. Rule one, every claim traces to a source. The system can’t override it, only I can.
Hooksthe enforcement
Automated gates fire on every action and block, warn, or pass. Rules with no enforcement quietly drift; hooks make the policy non-optional.
Playbooksthe learning
Named, repeatable procedures for recurring work. What succeeds is written down and reused; what fails corrects the playbook. The team gets better on purpose.
Agentsthe team
Specialised workers, each with a role and a scope. A researcher, a builder, a critic, a verifier, each takes a brief, reports back, and answers to the same rules. I direct; I don’t do it all by hand.
Where it pays off
Two ways it actually earns its keep.
The first is vertical. A thin feature bolted across everything rarely moves the number. Where AI compounds is depth inside one domain, owning the full flow end to end, where the context and accountability already live — well thought out, tested, and governed.
The second is quieter, and most companies miss it: running AI locally and privately, on their own machines. No data leaving the building, no per-token bill, full control. That’s how every company should run its system at scale.
The prompt isn’t the work. The governance is.
Rules it can’t override, enforcement it can’t skip, and the judgment of where to point it. That’s what turns AI from a gimmick into something that compounds.
Let’s talk