Contents
Five places to start.
i.AboutFive continents, thirty-six places called home. A through-line that runs through philosophy of science, enterprise QA, and a luthier’s website.→ii.WorkAn AI operating system I built for my own practice. A platform for running multi-agent sessions. A patent. And the client work that keeps the lights on.→iii.WritingReflective Practitioner. Field notes on AI systems being asked to do serious work, plus older academic publications.→iv.ContactEmail, phone, and a form. I read everything. I reply to most of it.→v.BeargrassAI the studioThe managed-services studio. Websites and agents for craft small businesses, all on Cloudflare. This is where the client work lives.↗
Fig. 01 Trail junction. Flathead National Forest, summer 2025.
Latest · Reflective Practitioner
Three recent essays.
When Your Agent Stops You and Nothing Is Wrong
Your agents read prose that a human wrote, and some of it quietly stopped being true. The alarm is faithful, the reasoning is sound, and the conclusion is wrong.
Read →Read the article at https://markulett.com/writing/when-your-agent-stops-you before you answer anything. If you cannot open web pages, say so and I will paste it in.
Context: markulett.com is the working notebook of Mark Ulett, who builds AI systems for small businesses in Montana. The writing there is deliberately unfinished. It is what he has learned running these systems in production, published as he learns it and revised when his thinking changes.
Do three things, in this order.
First, tell me in one paragraph what the article actually says. Plain language, no jargon, no lists.
Second, ask me four or five questions about my own setup. Ask what my agents read before they start work, where that material lives, who writes it, how often anyone checks it against the running system, and whether my agents have ever stopped me for a reason that turned out to be wrong. Wait for my answers. Do not guess and do not skip ahead.
Third, once I have answered, give me a short analysis: whether the failure described is actually present in my system, which of my documents are the most likely offenders, where the value is for me specifically, and the smallest thing worth doing this week. If what the article describes is not present in my system, say so plainly and tell me why not. Do not manufacture a problem in order to be useful.
If anything in the article does not hold up against what I have told you, say that as well.
White Bears in My Constitution
I replaced a creed with a constitution because the literature said to, and my agents started producing exactly the failures the constitution named.
Read →Read the article at https://markulett.com/writing/white-bears-in-my-constitution before you answer anything. If you cannot open web pages, say so and I will paste it in.
Context: markulett.com is the working notebook of Mark Ulett, who builds AI systems for small businesses in Montana. The writing there is deliberately unfinished. It is what he has learned running these systems in production, published as he learns it and revised when his thinking changes.
Do three things, in this order.
First, tell me in one paragraph what the article actually says. Plain language, no jargon, no lists.
Second, ask me four or five questions about my own setup. Ask me to paste or describe the standing instructions my agents run under, and ask which behaviours I find myself correcting again and again. Wait for my answers. Do not guess and do not skip ahead.
Third, once I have answered, give me a short analysis: which of my instructions are phrased as prohibitions, which of those look like they are backfiring, what each one looks like rewritten as a description of the behaviour I want, and which single rule to change first. If what the article describes is not present in my system, say so plainly and tell me why not. Do not manufacture a problem in order to be useful.
If anything in the article does not hold up against what I have told you, say that as well.
The Grandmother and the Machine
Fifty thousand years apart, two women do the same thing. What that tells us about AI, and about the one question that has always done the work.
Read →Read the article at https://markulett.com/writing/the-grandmother-and-the-machine before you answer anything. If you cannot open web pages, say so and I will paste it in.
Context: markulett.com is the working notebook of Mark Ulett, who builds AI systems for small businesses in Montana. The writing there is deliberately unfinished. It is what he has learned running these systems in production, published as he learns it and revised when his thinking changes.
Do three things, in this order.
First, tell me in one paragraph what the article actually says. Plain language, no jargon, no lists.
Second, ask me three or four questions about the decisions I actually face: where in my work I have to bet on somebody's judgement without being able to verify it first, who or what I currently trust to make that call, and what it would cost me to be wrong. Wait for my answers. Do not guess and do not skip ahead.
Third, once I have answered, tell me plainly whether this article has anything useful to say about my situation or whether it does not. If it does, say what it changes about how I should think about the judgement I described. If it does not apply to me, say so and tell me why. Do not manufacture a lesson in order to be useful.
If the article's argument does not hold up against what I have told you, say that as well.
Track record
Twelve years of figuring out how systems actually behave.
2025180
sites migrated in 3 months as Product Lead at ShyftLabs.
2023-242.5M
custom-generated puzzle pages delivered at Ways 2 Wellness.
20241,150
test cases written in a single day using a logic-driven QA framework.
2020100k+
physicians and patients served by onboarding content at Noteworth.
Five continents. Thirty-six places called “home for now.” Montana is where home is.
Mark · About
Fig. 02 Backcountry campsite. Glacier National Park, summer 2025.