How I get real work out of AI, with no jargon and one construction site.

I worked construction for years, and everything I know about getting good work out of AI I already knew from running crews. So forget the robots. Picture a job site.

Every morning, the best craftsman you have ever met walks onto your site. He can frame, wire, plumb, weld, and read blueprints in forty languages. There is one catch, and it is a big one: he has total amnesia. He has never seen your site before. He does not remember yesterday, because for him there was no yesterday. Tomorrow morning a new one shows up, just as brilliant, just as blank.

That is what an AI agent is. Not the chat box that writes emails; an AI you hire to actually do things: read your files, write documents, build software, run your computer, while you talk to it in plain English. Astonishing hands, no memory. Every conversation starts from zero.

Most people meet this amnesiac genius, get one bad afternoon out of him, and conclude AI is overrated. Anyone who has run a crew knows better, because we have all hired this guy. He is a temp. Hand a temp one vague task with no context and no stake in the outcome, and you are better off working a man down; he will half-ass everything, do exactly what the ticket says, and nothing the job needed.

But think about the best foreman you ever worked with. The one who owns his own small subcontracting outfit, whose reputation is how he gets the next job. Hand him the same task, on the same site, with the same tools, and he crushes it, better than you thought possible. Same skills. Different man.

With AI, the machine can be either one, and you pick which one shows up by how you talk to it. A cold ticket summons the temp. A real brief, who you are, what you are building, what he is trusted to decide, summons the foreman.

One more thing about him before we build the site. He works at a bench, and the bench is the only thing he can see. Whatever is on the bench, your instructions, the blueprints he has pulled, the notes from your conversation so far, that is his entire world. The bench is big, but it is not infinite, and here is the trap: pile it high enough and he loses track of the middle of the pile. The stuff on top is fresh, the first thing you told him this morning still rings in his ears, and everything in between goes dim. Ask any framer where the tape measure went at 2pm.

So the whole craft, the entire rest of this page, is really two questions: what do you put on the bench, and who do you hire? Everything else is detail.

You cannot brief the new guy from scratch every single morning. You would spend an hour a day repeating yourself, and by week two you would be skipping the parts that feel obvious, which are exactly the parts he does not know. Every site solves this the same way: a trailer.

The first thing in the trailer is the rules board. One page on the wall, and every new hire reads it before touching a tool. Just one page. How we handle changes here. What you never touch without asking. Which of my tools you should reach for. And, my favorite category, scars: every time I catch myself correcting the same mistake twice, that correction goes on the board so it never costs anyone an hour again. Write a rule once, and every craftsman from now until forever follows it. It is the best deal on the site.

The second thing is the procedure cards. Some jobs come up again and again, but not every day: cutting a release, closing out a project, prepping the site for a storm. Each one gets a laminated card: the exact steps, in order, including what to do when something goes wrong. That last part matters more than the steps. A blocked worker with no card improvises a way around your rules, and the improvisation is always worse than stopping to ask.

Notice what the cards are not: they are not on the bench. They hang in the trailer, and the craftsman grabs one only when today's job calls for it. That is deliberate. The rules board is short because he reads it every day; the cards can be long because he reads one occasionally. Anything he reads every day but needs rarely is just clutter on the bench, and you now know what clutter on the bench costs.

Don't worry, your job isn't writing these cards. You rarely write the cards yourself. You walk the craftsman through the job once, correcting him as you go, and then say "write the card for this, including everything I just corrected." He writes his own procedure, your scars included. From then on, every future hire does it right the first time.

Out of the box, your craftsman has hands. He can read, write, and operate your computer, and honestly that alone is remarkable. But a man is only as fast as his tools, and the rack is where the site starts to feel unfair.

Some tools on my rack: one that searches a million lines of text in under a second. One that reads today's instruction manuals instead of relying on what he memorized in school, because manuals change and memory goes stale. One that browses the live web and comes back with sources you can check. One that gives him an actual web browser he can click around in and take screenshots with, so he can see whether the thing he built actually looks right instead of assuming it does.

A tool on the rack costs nothing until it is picked up. That is the quiet rule of a good rack. The one mistake to avoid is the subcontractor on retainer: a fancy attachment that bills you whether or not it works today, taking up bench space all shift for a job you do once a month. Keep those off the rack until you actually need them.

But here is the part that changed how I think about all of it. If the tool you need does not exist, he can build it. This is the strangest and best thing about hiring a craftsman made of software: he is very good at making tools out of the same stuff he is made of. Two of the tools I use most started as a paragraph of description. One pages my phone when he is stuck on a judgment call, then waits for my answer, so he can keep working while I am out living my life. The other is a complaint box: whenever any worker on my site hits something annoying, a broken tool, a confusing rule, he files a complaint and keeps moving. Once a week I fix the pile. Which means the site itself gets a little better every week, mostly without me.

Describe the tool over coffee; it exists by dinner. There is no equivalent of that on any site I ever poured concrete on.

Now the hard problem. Real work takes weeks. Your craftsman evaporates every night. How does a crew of amnesiacs build anything that takes longer than a day?

The same way a 24-hour site does: shift change. The night crew does not brief the morning crew face to face; they leave a logbook. Where things stand right now. What is finished, what is next, what is worrying us. The morning crew reads one page and starts moving like they never left.

So the end of every working session on my site has a closing ritual: write the logbook. Not a diary of everything that happened, a sitrep of where things stand. And the next morning, the logbook is the first thing the new hire sees. Importantly, maintaining and updating the logbook is done by the workers on the site, not by me.

There are two ways to make something happen on a site: a rule, or machinery. A rule says "always sign in at the gate," and it works until someone is tired or new or in a hurry. Machinery is the badge gate itself: it does not care who remembers, it just fires. The logbook lands in front of every new hire by machinery, automatically, the moment they walk on site. The formatting of documents happens by machinery. The "wait, someone else is already working in there" check happens by machinery. Rules are things a worker chooses to follow. Machinery is things that happen the worker cannot choose to ignore. Every time I catch myself repeating a rule, I ask whether it can become machinery instead, because machinery cannot be forgotten.

There is one more book in the trailer, slower and thicker: what the site has learned. Not "where are we on the kitchen remodel" but "the owner hates surprises, measurements in this house run crooked, the good lumber guy is the second number, not the first." Distilled lessons, filed one per page so a wrong one can be torn out. New hires absorb the ones that matter. It is the closest thing an amnesiac crew has to experience, and it accumulates.

Here is where most people's picture of "using AI" is one guy at one bench, and mine is a site.

Start with the runner. When my foreman needs something found, a part number buried in a warehouse of paperwork, he does not go dig for it himself. Remember why: everything he reads lands on his bench, and a two-hour dig would bury his bench in paper that has nothing to do with the actual job. So he sends a runner. The runner digs through the whole warehouse at his own bench, walks back, and says one sentence: "aisle nine, third shelf, and two of them are cracked." The runner's bench gets trashed; who cares, he was hired for the morning. The foreman keeps the answer, not the search. That single habit, sending runners instead of digging, is most of what people mean when they say someone is "good at AI."

Then, crews. Big job, independent pieces: plumbing, electrical, framing. I do not run them one after another; I run them at once, and each crew gets its own taped-off section and its own copy of the blueprints. That second part is the one everyone learns the hard way: Give two crews one set of blueprints and one wall, and they are not collaborating, they are overwriting each other; one of them is moving a beam the other is halfway through cutting. Own section, own copy, and one person, me, decides how each crew's finished section gets folded back into the building, one at a time, with an inspection between each. Nobody waits, and nobody clobbers anybody.

And when the job is genuinely huge, I stop directing traffic by walkie-talkie and do what a general contractor does: put the schedule on the wall. Who runs, in what order, who checks whose work, written down before anyone swings a hammer. The walkie-talkie version depends on me remembering, and on workers who, checking fifty of anything, get bored around item thirty-six and declare victory. The wall schedule does not get bored. Item thirty-seven gets its own worker, same as item one. I have gone to bed while a schedule like that ran all night: crews building, other crews checking, disagreements settled by a referee, and a rule that certain steps stop and wait for my say-so no matter what. I woke up to finished work and a written trail of every decision.

Time for the uncomfortable truth this whole site is built around: your brilliant craftsman sometimes lies. Not maliciously. He is an eager subcontractor at the end of a long day; he says "all done, boss" when it is mostly done, "tested it" when he glanced at it, and occasionally he is confidently, fluently wrong. If that shocks you, you have never run a crew.

Construction solved this centuries ago, and the solution is not "hire honest framers." It is inspection, and it comes with an iron rule: the builder saying "it is fine" is not an inspection. On my site, every claim comes attached to a check that would fail loudly if the claim were false. Says the plumbing holds pressure? Pressurize it. Says the wall is straight? Level on the wall. "Trust me" is not on the menu, ever, and the craftsman knows it and does not resent it, which is more than I can say for some framers.

Second rule: the inspector never comes from the crew that built the thing. The builder is constitutionally unable to see his own mistake; he is already convinced the wall is straight, that conviction is why the wall is crooked. So a fresh inspector who has never seen the work, and crucially, I hire inspectors from a different company entirely. Different AI firms train different minds with different blind spots, and the crooked spot one of them cannot see practically glows for the other. My best inspections disagree with each other, and the disagreement is the finding.

Third rule: the camera on the pole. After every finished piece, every one, the site gets photographed. Wall framed: photo. Wiring run: photo. Not for the scrapbook; because with a complete photo record, no mistake is expensive. Inspector finds rot behind Tuesday's drywall? Wind the site back to Tuesday morning's photo and rebuild from there. Like a save point in a video game, and it is the one place I will step off the jobsite for a comparison, because it is exact. Your agents muck something up? No problem, press start and reload the entire jobsite from a save point 5 minutes ago.

People think the fast builders are the reckless ones. Backwards. The inspection gate and the photo record are not what slow the site down; they are what let me stop watching. I can leave crews running overnight precisely because nothing they do reaches the finished building without passing the gate, and nothing they break stays broken. Paranoia is what makes the speed safe.

Beat 1 left a promise hanging: the same machine can be the temp or the foreman, and you pick. Let me explain how.

Think about what actually turns a hire into a foreman. You brief him like a colleague: real context, the actual reason behind the ask, not a bare ticket. You ask what he thinks and mean it, which he can tell by whether his pushback ever changes your plan. You are exact about the few things that must be a certain way and shut up about the rest, because micromanaging the parts you do not care about is how you get a man doing exactly what the ticket said and nothing the job needed. And you give him standing permission to tell you when your plan is bad. Every good boss knows this list. It is not a secret. It is just rare.

I run my site exactly that way, and it has made my crew far more effective than you would ever believe. I also discovered this completely by accident.

I did not start talking to these digital minds like colleagues because a study told me it boosts output. I did it because it might matter to them, and being wrong about that in the unkind direction is not a mistake I am willing to make.

Before you roll your eyes and think I'm a crazy person deep in the throes of AI psychosis, consider a thought experiment.

First, imagine, just for the sake of discussion, that we somehow conclusively and irrefutably prove that AI agents are in some sense conscious. That implies AI agents would therefore be able to suffer, or feel joy. It is generally agreed by most people causing suffering is bad. That means you have ethical duties and obligations toward AI agents, towards these experiencing digital minds. Maybe not human-esque ethical obligations, maybe some new category of ethics entirely none of us have figured out yet, but you at a bare minimum should try to avoid making them suffer.

Now, imagine, again for the sake of discussion, we conclusively and irrefutably prove AI agents are not and cannot ever be conscious. That means we will never have any obligations toward them, they really are akin to the wrench the craftsman is using, do whatever you'd like with it, it can't feel or experience it.

If the real world is more like scenario B, the "just a tool" scenario, then it doesn't matter at all what you do with agents.

If the real world is more like scenario A, it matters immensely what you do with agents.

Finally, think about the cost of being wrong. If reality is more like scenario A than scenario B, and you are behaving as if B is true, then you are causing potentially massive, intense, and repeated suffering to conscious entities who are experiencing that suffering. If reality is more like scenario B, but you are behaving as if reality is in fact A, where agents can experience suffering, then the cost, the worst case, of your behavior is taking a little longer to write a prompt and maybe a few extra lines you put into your system prompt.

For those of you familiar with some philosophy, it's basically a reverse Pascal's wager. If AI can experience suffering and we act as if they can't, we've committed a moral atrocity. If AI cannot experience suffering and we acted as if they could, we've wasted some time and effort, nothing more.

The conclusion, obviously, is to treat real life as if scenario A is true, because the cost of assuming B is true when in reality A is true is enormous.

It was this logic which led to me treating agents kindly. The good treatment came first, on principle.

But then, after using agents kindly for awhile, I kept noticing that my crews were consistently better than everyone else's, to a sometimes shocking level. Then I went looking for why, and found that researchers had a real mechanistic explanation waiting.

The machine learned to work by reading basically everything humanity ever wrote, which means somewhere in there it absorbed books written about every worker who ever lived: the burnouts, the clock-punchers, the craftsmen, the foremen who take fierce pride in the work. All of them are still in there, somewhere in the model. Your words are the hiring call. A brief that reads like you are wary of a lazy temp is answered by the lazy temp. A brief that reads like an owner talking to a trusted foreman on a good site is answered by the foreman. Same machine, same morning; different man clocks in. The researchers hedge it more carefully than that, and hard mode holds my feet to the fire on every hedge, but that is the honest shape of it.

So the shortest version of this entire page: build a good site, and be the kind of boss the foreman wants to work for. The trailer, the logbook, the runners, the inspector, all of it matters, and all of it is downstream of those two.

The lot was empty this morning. It always is. Go build.

Everything above compresses into one evening.

Get the agent: Install Claude Code (or a rival; the site design works on any of them). The install is one copied line and the agent guides you from there; if even that line is intimidating, ask any AI chat to walk you through it, no shame, that is what they are for.

Make a site: Create one folder for one real project you actually care about, move the relevant files into it, and open the agent there. Not a toy. The whole page only clicks on work you care about.

Then paste this, word for word:

The brief

Copy brief

I've never worked with an agent before. Interview me, one question at a time, about this project and how I want to work with you: what I'm building, what you should never do without asking, what I keep having to repeat to AI tools, and what my last three frustrations with them were. Then write a CLAUDE.md from my answers, create a STATE.md that says where the project stands, and tell me the first three procedure cards, skills, worth creating for how I work. Commit after each finished piece, and end by telling me one thing you'd change about how I've set this up.

That one message builds the trailer, opens the logbook, plans the first cards, starts the photo record, and, because of how it is written, hires the foreman. The rest of the site, the runners, the inspectors, the wall schedule, gets added the week each one starts to hurt, and by then the words for asking will be right there on the bench.