What the eight weeks actually cover
Week 1 is what these tools are, taught through the phone keyboard's next word suggestion, and the two sentences every attendee has to be able to say cold. Week 2 is the difference between something that talks and something that acts, which is where the course delivers on its own name. Weeks 3 and 4 put a website and then a working web app on the internet during the session. Week 5 is answers from your own documents. Week 6 turns the specification written in Week 2 into a running agent. Week 7 is models on your own machine with nothing leaving the building. Week 8 is a three minute demo from everyone, no exceptions.
The arc was deliberately rewritten so first contact with an agent moves to Week 2 rather than Week 6, and prompting stops being its own week: context, constraints and the ask get taught inside the hands-on blocks and sharpened every week after.
Six teaching rules, applied to every section
Analogy first, term second, so no technical word appears before a concrete picture of it. Every term gets a plain English definition in the same sentence it first appears. Attendees spend more of the session doing than listening. Each week has anchors, short sentences repeated at least three times, word for word, never varied. Each week builds a fence: an explicit list of what the thing is not. Verification is taught as a physical habit rather than a disclaimer.
That last rule is the course in miniature. The verification habit is lifted straight from electrical work: nobody trusts that a circuit is dead because somebody said it is dead, you lock it out, tag it out, and check with your own meter. The anchor repeated all night is "confidence is not accuracy."
The demo that does the teaching
Week 1 runs a fabrication on purpose. Search is switched off and the room is asked for the meaning of fault code E-47 on a compressor model that does not exist. The answer comes back confident, specific and numbered, and entirely invented. Then the same question runs with a real manual pasted in, and the answer changes and cites.
The line delivered after the reveal is the point of the whole evening: the machine did not get smarter between those two answers, it got material. Everything the course teaches afterwards, retrieval, your own documents, local models, is a way of giving it material.
In the code
/* Pity timer: any active Build List goal whose tag has not resolved a card in 4+ days. */
pityTags: function () {
var r = NG_STATE.run, out = [];
this.milestones().forEach(function (m) {
if (m.done || !m.tag) return;
var last = r.lastTagDay[m.tag] || 0;
if (r.day - last >= 4) out.push(m.tag);
});
return out;
},
drawWeight: function (card, pity) {
var w = card.weight;
var tags = card.tags || [];
for (var i = 0; i < tags.length; i++) if (pity.indexOf(tags[i]) >= 0) { w *= 5; break; }
return w;
},The course ships with a 56 day narrative game covering the same eight weeks, and this is its honesty mechanism. A pure random draw lets a run stall: the goal you are working toward simply stops coming up, which is a software artefact, not a lesson. The pity timer multiplies the weight of any goal that has gone four days without progress, and a separate crisis floor swaps the deck for a recovery deck when momentum collapses, so a bad run bends rather than ends. Both rules exist because the failure they prevent is the one that makes a learner quit, and the course is aimed at people whose first instinct on a stall is to conclude the fault is theirs.
How this differs from the ordinary version
Owning a small system beats renting a large one
Third party software is built for the average of thousands of customers, which means every business pays for the parts it does not use and works around the parts it needs and did not get. The course teaches the other path: describe your own workflow, build the narrow thing that runs it, keep the data. By Week 6 an attendee is holding a specification they wrote in Week 2 and an agent that runs it.
The bill stops being a subscription
A rented tool charges per seat, forever, and the price is set by the vendor. Something you built costs what it costs to run, and Week 7 exists specifically to show that a capable model can run on hardware already sitting in the building, at which point the recurring cost of the thing approaches the electricity.
Your data does not leave
Handing customer lists, medical records or contracts to a hosted tool is a decision most small operations make without noticing. The course names that in Week 1 as a failure mode and answers it in Week 7 rather than leaving it as a warning.
It is free, local, and ends in a demo
Eight weeks, no cost, in El Paso, and the last session is everybody presenting for three minutes. That last constraint is deliberate: a course that ends in a certificate produces certificates, and a course that ends in a demo produces working things.
In the field
A worked example the room can open
ATRIUM is the argument in finished form: a workflow intelligence platform built for one specialty medical practice, architected for its record handling requirements, rather than a general practice management suite the practice would have bent itself around. It is the answer to the question every attendee eventually asks, which is whether the thing they build can be real.
Why this region
El Paso, Santa Teresa and Ciudad Juarez run on small operations and family shops that a per seat subscription prices out and a national vendor never visits. The people in the room already understand procedures, tolerances and verification from their day jobs. The course borrows that vocabulary rather than replacing it.
Questions
- What does it cost?
- Nothing. It is a free community course run in El Paso, and every tool used in the first weeks has a working free tier.
- Do I need to know how to code?
- No. Week 1 starts at the phone keyboard's next word suggestion and every technical term is defined in plain English in the sentence where it first appears. More of each session is spent doing than listening.
- Why build instead of buying software that already exists?
- Because a general product is the average of thousands of customers, so you pay for what you do not use and work around what you needed and did not get. A narrow system you own runs your workflow, keeps your data, and costs what it costs to run rather than a per seat fee set by somebody else.
- What do I leave with?
- A working account, three completed real tasks from the first night, a specification you wrote in Week 2, an agent built from it in Week 6, a model running on your own machine in Week 7, and a three minute demo of your own work in Week 8.