The Semester Contract: Your Business Is the Case Study
Picture the last week of the semester. On the screen is a live page where anyone in the world can buy something you made, not a mockup in a slide deck, a link that takes money. At least ten people have used it, and at most five of them are classmates, which means at least five strangers decided your thing was worth paying for. Beside it is a ratio: what one production run of your product costs you now, against what it cost you in minutes on the first Thursday of the semester. Both ends measured by you. In your hand is a journal with seventeen weeks of dated entries: decisions, predictions, the ones you got wrong, and the measurements that proved it.
That's demo day, and every single person in this course can reach it. The constraint isn't talent and it isn't capital. Nobody incorporates anything, nobody buys ads, nobody needs a co-founder. The constraint is whether you actually start today, and whether you measure honestly enough to steer.
This is the second recorded unit of Class 1, building directly on Unit 1's "why now" argument. If you haven't written down a real recurring task from your own work yet, do that first. This unit assumes you have one in front of you.
Why your own business, and not a case packet
Every case in this course has the same protagonist: you. There's no case packet from a distant business school, no disguised company with a tidy dilemma solvable in forty minutes. The case method is a good method, and this isn't a dismissal of it, but it teaches one specific skill: analysis under someone else's framing, with the irrelevant material already stripped out by an author who knew the answer. That's useless training for the skill this course actually exists to build, which is steering your own operation when nobody has told you which facts matter.
There's a structural reason specific to this subject, too. A zero-employee organization is steered by one person's judgment against one person's numbers. The whole discipline is a single accountable human reading their own instruments. Studying someone else's instruments teaches you commentary. Studying your own teaches you operation. In a course whose founding rule is clocked, not felt, the only defensible data source is the clock you're holding yourself.
The shippability test
Before choosing anything, here's the single most reusable framework in this class, and it outlives the course entirely: five questions for evaluating any product idea in an agentic era.
Can an agent produce most of the artifact? Not all of it, most of it. If the deliverable is text, structure, analysis, layout, code, or a document, the answer is usually yes. If it requires physical presence, a license you don't hold, or a relationship that must be personally carried, the answer is no, and that's a fine business, just not this semester's.
Can you tell good from bad, fast? This is the one people skip, and it's the one that kills them. If you can't look at the output and judge it in a couple of minutes, you can't ship it, because you'd be shipping something you can't stand behind. Remember the governing sentence from the last unit: capable but not accountable, and accountability is non-transferable. Your competence as a verifier is the actual ceiling on what you can sell. Pick a domain where you're a competent verifier, which usually means picking something you already know.
Does it deliver as a file or a link? Anything requiring logistics, inventory, or a physical hand-off is out of scope by the time-box. This isn't a judgment about business quality. It's a judgment about what fits in a Thursday lab.
If it's wrong, what breaks? Agents are confidently wrong. Choose a product where being wrong is embarrassing and correctable, not harmful. Nothing that constitutes medical, legal, or financial advice to a specific person. Nothing where a hallucination lands on somebody who trusted you. This is the same logic as the discount limit the last unit specified for Claudius (the AI instance Anthropic put in charge of a real shop in the Project Vend case), not squeamishness.
Is every input yours? Every input has to be your own material, your own knowledge, or properly licensed. Nothing that trades in anyone else's exams, question banks, copyrighted material, brand, or datasets. That fence is absolute, and it's the one term of this contract with no negotiable edge. A product that fails this test can't be graded, can't be sold, and can't be defended, and the semester you spend on it is a semester you lose entirely.
Run any idea through those five. Four yeses and one no means change the idea, not the rule.
The six shapes
To make choosing fast instead of paralyzing, here's a catalog of six shapes that pass all five tests. Notice the pattern as you read: each one is distribution you already own, plus judgment you already have, plus bulk an agent can produce. The shapes are just instances of that same pattern.
| Shape | What it is | Why it clears the tests |
|---|---|---|
| Recruiting-season packs | Interview question banks by sector, company one-pagers, application trackers, timeline calendars | You're inside the hiring cycle right now, which makes you a competent verifier, and classmates are reachable buyers |
| Study systems from your own material | Your notes restructured into summaries, practice sets, spaced-repetition decks | The ownership test governs absolutely: your material, your questions, never anyone else's exams |
| Onboarding packs for arriving exchange students | Housing, banking, phone lines, paperwork, neighborhoods, what nobody tells you in week one | A new cohort needs exactly this every semester, and the people who know are busy while the people who are free don't know |
| Micro-guides to a niche you verifiably know | Something specific you can defend under questioning | The narrower the better, specificity is both the moat and the proof of your competence |
| Professional templates | Financial models, thesis setups, pitch skeletons, research protocols | High perceived value, agent-producible bulk, easy verdict test if it's your own discipline |
| A mini-course teaching a skill you have | Structured lessons, a worked example, an exercise | Highest effort, highest price, strongest signal, and the shape most likely to outlive this class |
You may propose something off-catalog. The only disqualifier is the time-box: your operation must fit inside the Thursday labs plus at most two hours a week. This is one class among several you're taking. A product that needs more than that is out of scope by definition, and choosing inside a constraint is the first act of operating, not a compromise of it.
Why one stranger, not ten customers
Be precise about the bar, because misreading it means optimizing the wrong thing all semester, or worse, spending the semester anxious.
Required and graded: shipped, offered, measured. A working payment link, actually put in front of real buyers, with the funnel numbers in your journal. This part doesn't depend on luck or a market cooperating, which is exactly why it's the part that's graded.
Zero is the only real failure, and it means you never actually offered it. That's entirely within your control, which is why it's the one thing worth pushing on. One stranger is the thing that can't be faked: someone with no social obligation to you looked at it and paid. Effort can't manufacture that fact, and neither can a good deck. Three means it repeats rather than happened, that's the target. Ten is the stretch, and it's the instructor's number, not yours, deliberately harder because more practice and more hours are already banked.
You'll sell at small prices, as an individual, through a payment link. Nobody incorporates anything. The money is trivial by design, because price isn't the variable under study this semester. The minutes are.
If you reach one customer and your journal shows the funnel, what you tried, what each attempt cost in minutes, and what you changed in response, that's a semester of real operating and it grades as one. An unmeasured excuse is not.
Milpapel: built alongside you, in public
This semester's instructor is building Milpapel, an online print shop, from nothing, in front of the class. Every technique in this course gets applied to it the same week it's taught: the actual commands, the actual failures, the actual numbers, including the instructor's own baseline card and ratios, subject to the same rules as everyone else's. The target is ten orders by December. If Milpapel works, the class has a worked example running a week ahead of them all semester. If it doesn't, they get something more useful: a diagnosis, in public, by the person who said the method works.
The terms of the contract
Before leaving this unit, you'll have a working name, a shape, a buyer, and the workflow that produces it. Not a plan to decide later, today. What's not accepted after today is the toy: a product you'd never actually sell. Toys have no stakes, and without stakes every technique in this course collapses into trivia.
The labs produce telemetry: your minutes, your token costs, your workflows. Every numbers-moment in the Controlling block, every variance analysis, every steering decision this course practices, runs on your baseline and your telemetry, not on figures invented to make the arithmetic clean. This is the flipped deal of the whole course: the instructor supplies a century of method, you supply the data it runs on. Neither works alone.
Every figure leaves your notebook only as a ratio to your own Week One baseline: "this workflow now costs 0.4 times my week-one minutes," never pesos, never dollars, never wages or totals. Your denominator stays private; only the multiple travels. The cohort can compare progress, whose workflow costs collapsed, whose plateaued, who found a method worth copying, while nobody discloses a single absolute financial fact. It's the same move economics makes with index numbers: strip the units, keep the movement.
Dated entries, from today to demo day. Decisions with dates. Predictions with dates. Measurements with dates. Revisions with reasons. The journal is graded for evidence of steering, never for the success of the venture. A student whose venture struggles all semester, documented honestly, measured carefully, steered visibly, outscores a student whose venture thrived while their journal shows nothing but victory laps.
A measured failure banked in your journal outranks an unmeasured success story told in class. Self-reported feelings missed the direction of the effect in the randomized controlled trial from Unit 1's "Why Now" lesson (sixteen experienced developers, 246 real coding tasks). Clocked, not felt. Measured, then claimed. "I did not measure it" is a respectable sentence in this room. "It felt faster" is not.
The four ways to lose this semester
Naming these in week one prevents most of them.
| Failure mode | What it looks like | Antidote |
|---|---|---|
| The toy | You pick something you'd never sell, because it's safer | The buyer test: could you message someone tonight, without an introduction, who might actually pay? |
| The infinite build | Fourteen weeks perfecting the machine, no shipped product, build-setup hours hidden off the card | Build-setup minutes logged at full weight from today; the first sale is due long before you feel ready |
| The unmeasured claim | Real, solid work with nothing to prove it | From today, "done" means done and costed |
| The entanglement | Building on material that isn't yours | The ownership test, applied before you start, not after |
The infinite build is worth naming twice: it's the most common failure among the strongest students in the room, and it will be tempting in a real way, not a hypothetical one.
The working session
Take twenty minutes now, before moving to the lab in the next unit, and work through this in earnest.
Write down the one recurring task you're bringing into the semester as your case study.
Score it against the five questions above, honestly. Not the version of the task you wish you had, the real one.
Note which mode (by-hand, agent-assisted, agent-run) it's realistically a candidate for today, and why.
The next unit turns this into your first real measurement.
If you're following this course outside the live class, there's no room full of classmates to compare notes with in real time, but the exercise works exactly the same alone. Write it down as if you were about to explain your choice to someone else, because in the next unit, you effectively will. You're about to measure it.
Before you finalize your product, spend three minutes trying to fail it on the five tests yourself, specifically and quickly, the way a skeptical classmate would, and then find the nearest version that passes. If you're stuck between two shapes, pick the one whose buyer you could message tonight without an introduction. Access beats elegance. You can revise the product; you can't conjure an audience.
Your journal's first formal entry
Record your product's working name, its shape, its buyer, the workflow that produces it, and today's date. Then two predictions, one sentence each: by demo day, what ratio-to-baseline do you expect your production workflow's operator-minutes to reach? And in which week will you make your first sale to a stranger? Sign both.
You'll grade both predictions yourself in week seventeen, against your own measured number. Nobody else grades it, and it costs you nothing to be wrong. The point is to discover, seventeen weeks from now, what your calibration was worth before you had instruments.
Next unit: the instrument itself. Bring the workflow.
Reply here and it goes straight to Rod. Same as replying to one of his emails.
