How I work
Field notes on letting AI do the work
The rules I actually run on, at work and in my own business.
-
Set it up so AI can do the job
AI as a faster pair of hands got a 6-to-10-month build done in about 2 months. In 2026 I flipped it: I set up the environment, access and rules, and AI runs the project.
More on this: Set it up so AI can do the job
I'm asked for judgment at the start and at named decision points. The first time costs more than doing it yourself; call it the enablement tax. Judge the model on project two. Ours merged 129 changes in 26 days.
I no longer am using AI to accomplish projects. My goal is focusing on safe and secure AI enablement so that AI can do the project.
-
Context engineering, not prompt engineering
The prompt is the least important part. What matters is the environment: rules files at every level, reusable skills, connections to the real systems, and memory that survives the session.
More on this: Context engineering, not prompt engineering
Vague documents make a confidently vague AI. When my triage agent prescribed the wrong runbook, it had matched on the title alone. A sloppy document name used to be a nuisance. Once an AI reads your docs, it's a defect.
-
Trust is earned in stages
Every agent I build starts at stage one: it proposes, a person approves every action. Then a narrow lane, reporting everything. Then its full lane, reporting exceptions. Then a schedule, with a periodic audit.
More on this: Trust is earned in stages
What widens is the target, never the capability. My patching agent could do the same things on day thirty as on day one. What changed was what it could touch.
Promotion is a written decision, and so is demotion: when a batch left two machines mid-upgrade, I pulled the grant for cause.
Trust is not granted. It's a privilege, like it is with a person, and it's only offered based off of earned results.
-
Fence it with credentials and code, not promises
I don't trust an AI to stay in its lane because it said it would. I build the lane so it can't leave. Most of the effort goes into the fence.
More on this: Fence it with credentials and code, not promises
For the first agent allowed on infrastructure: a key bound to one address and one command, a seven-verb gatekeeper, a non-production allowlist, an audit log, and a cooldown only I can clear.
The AI itself argued it must not be able to clear that cooldown. I brought the fence to leadership with the ask; approval took about 90 seconds.
Scope is enforced by credential and code, never by empty AI agent promises. We don't trust it to stay in its lane. So we built it so that it can't leave its lane.
-
A sentence saying it worked is not evidence
Machine-checkable output or it didn't happen. Read-back is the verdict, never the return code. Check the machine before you trust the report about the machine.
More on this: A sentence saying it worked is not evidence
It catches silent failures, the real risk: green runs that skipped their work, alerts that never fired. Mine once didn't, for weeks: they posted as me, and Slack doesn't notify you of your own messages.
A second model reviews as an adversary before anything reaches a human; it has caught a policy ID that didn't exist. My manager's AI once caught an approval check that matched “I do not approve.”
-
Skills are jobs. Agents are workers.
A skill is a repeatable piece of work done the same way every time. An agent is the worker that runs it. Keeping those separate makes the work shareable.
More on this: Skills are jobs. Agents are workers.
If my manager or a teammate could do something I do, it goes in the shared repository and they get it with one clone.
My agent-builder method interviews me one question at a time, forces a one-sentence definition of done, and may answer “this should be a skill, not an agent.” Many ideas end there, which is the point.
-
Keep human judgment at the right place
It will make mistakes. Of course it will. The question is where a person sits. I keep judgment, approvals, scope calls and verification. I delegate execution, bounded design, and the digging.
More on this: Keep human judgment at the right place
Domain knowledge still beats the model, and you have to know when. I stopped an agent from running against production: it would have restarted a web server still taking traffic. It's an excellent employee. It isn't the company's God.
It's not about whether or not it's going to make mistakes. Of course it is. The point is keeping human intelligence at the right place.
-
Faster tools make focus worth more
AI raised my in-flow speed, which made every interruption more expensive: you're dropping out of a sprint, not a jog. The extra capacity should become better focused work, not more things crammed into an afternoon.
More on this: Faster tools make focus worth more
So saying no has to get harder as velocity rises. I block my mornings and measure the week by whether committed work moved.
Owners will feel this first: when the paperwork takes an hour instead of a night, the temptation is to fill the night.
-
Systems over heroics
Teams get pulled into reactive work by what I call operational gravity: every undocumented fix, every process in one person's head. The way out isn't saying no. It's the system that makes that work disappear.
More on this: Systems over heroics
That's why automation never lives on my laptop, skills go in a shared repository, and an agent watches our automations so nobody learns something broke by hitting its consequence.
Losing a person should never mean losing every system they built. The Navy taught me that first: I rebuilt the training program instead of tutoring, and advancement went from 8% to 85%.
Automation on one engineer's laptop fails exactly the disaster where that engineer is unavailable.
Where this stops
AI gets things wrong in ways that look right: a quote you never said, the wrong root cause chased with confidence, success reported on work it skipped. Everything above exists because of that.
My own limits: I over-scope, and I've learned to cut go-lives down to hold a date. Bare-metal hardware and low-level networking are my weaker lane; cloud is where I'm strong.
I direct AI to write automation, but I'm not a developer and won't sell you an app.
