Words you'll hear
The syllabus uses plain words. Canvas, the log form, and older slides still use some of the original ones. Here's the translation.
- Your assistant (also: your agent)
- The AI you build and run on your laptop with Claude Code. It's yours alone, even though your project is shared.
- The pieces of your assistant
- A skill is a saved instruction you can run by name (like
/card). A hook is something that runs by itself when an event happens (a file gets saved, a command is about to run). A rule is a standing instruction it always follows. Memory is what it remembers between sessions. A helper agent is a smaller assistant with one job. Together, the pieces are what the original syllabus calls "the harness." - A tedious job (the original says "inefficiency")
- Work a real person does by hand, the same way, every week. Scheduling, reminders, re-typing information from one place into another, chasing people for things, looking things up. Your project fixes one of these.
- Your partner business
- The one real organization where your assistant will actually run. A business, nonprofit, practice, farm, shop, or firm. Your dorm and your clubs don't count.
- Your contact (the original says "process owner")
- The person at your partner business who does the tedious job today. They agree to meet your team three times: once to explain the job, once to try your first version, once when it's running for real.
- The other ten
- Ten more organizations that do the same job the same way. Your partner business is where you build it. The other ten are why it's a business and not a favor.
- The three steps (the original says "touch points")
- Every tedious job has the same shape: information comes in → someone decides something → information goes out. Your assistant helps at each of the three steps. At each one it can collect the information, summarize it, flag what matters, or draft what goes out. A person keeps the decision until you can write a rule that makes it safe to hand over.
- The two numbers
- Two numbers I'll ask your team for every week from Week 3 on: how many real people you've talked to so far, and how many of the three steps your assistant handles for real (with real information from your partner business, not made-up examples).
- Satisfactory / Needs Revision
- The only two marks any assignment gets. Satisfactory means it meets the checklist. Needs Revision means fix it and turn it in again. No points, no percentages.
- Tokens
- Two passes you can spend all term. Each one buys a 72-hour extension, a redo of something marked Needs Revision, or one forgiven absence.
Finding your tedious job
You already turned in three candidates and put them on the wall. Your team picks one to start with; the team plan due Sep 25 says which. Here's how to tell a good one.
Ten questions to ask your contact
Ask these in your first conversation. Write the answers down word for word. The best target usually shows up in questions 3, 5, and 8.
- What do you do every week that you could teach a new hire in an afternoon?
- What information do you go find before you can make a routine decision, and where does it live?
- Which task do you put off because it's tedious, so it piles up?
- When you say "let me check and get back to you," what are you checking?
- What gets copied from one place into another by hand? Email to spreadsheet, invoice to QuickBooks, form to calendar.
- What decision do you make by the same rule every time, but still have to be the one to make?
- Who waits on whom? Where does a job sit until someone remembers to hand it off?
- What did you not follow up on last month that cost you money?
- If you were out for two weeks, what would break first?
- What do you write that's 80% the same as the last time you wrote it?
Four tests before your team commits
- Could they write the steps down? If your contact can dictate a checklist for the job in ten minutes, an assistant can do it. If they can't, you probably picked a judgment call.
- How long does it take now? Minutes per week, how often it's late, how often it's wrong. No number, no project.
- Can AI do it at all? Try the job with a plain chat prompt on three real examples before you build anything. If it's bad at the easy cases, pick something else.
- Would ten other businesses pay to stop doing it? Name them. If you can't name ten, you've found an errand.
What a good target looks like
The automations working in 2025 and 2026 share three traits: the task is narrow and eats a lot of hours, the rules for doing it right can be written down, and a person still checks anything that gets cited, invoiced, or promised. The ones failing need judgment in unfamiliar situations, trade on trust, or need facts exactly right with nobody checking. Legal research, medical advice, and anything that pretends to be the owner are off the table.
Developers in one 2025 study felt 20% faster with AI and measured 19% slower. AI legal-research tools still get one citation in five wrong. An airline was held liable for what its chatbot promised a customer. That's why you measure first, why a person stays at the step where facts must be exact, and why "the AI did it" is never a defense.
Examples, by kind of business
Each line is: the tedious job → what your assistant does → what the person still does. These show how narrow "narrow" is; pick your own.
- Restaurants and taprooms. Reviews piling up on Google, Yelp, and delivery apps → collect them, draft replies in the house voice, flag anything about illness or staff → owner approves every reply. Weekly reorder by gut → pull sales, inventory, and supplier prices, propose the order with reasons → chef edits the quantities. Brewery compliance filings → gather the logs, prepare the forms → brewer signs.
- Contractors and trades. Estimates that take days → site notes and photos in, the firm's template filled from their price list → contractor sets the margin and signs. Missed permit deadlines → watch the county portal and the job calendar, draft submissions, remind → office manager submits.
- Retail and repair shops. "Is my repair ready?" calls → read the ticket status, text customers when it changes → technician updates the status; price disputes go to a person. Seasonal pricing → scan competitor prices weekly, flag overstock, propose markdowns inside limits the owner set → owner approves.
- Farms and CSAs. Weekly member email → gather the harvest sheet, draft the newsletter, update the pickup list → farmer edits and sends. Grant monitoring → scan opportunities, score the fit, draft summaries → farmer decides which to pursue.
- Small nonprofits. Donor thank-you letters → pull gifts from the database, draft the letters → director signs; big gifts flagged for a personal call. Grant reports → assemble outcomes and expenses into the funder's template → program director checks every number. Volunteer scheduling → match sign-ups to shifts, send reminders, fill no-shows from a waitlist → coordinator resolves conflicts.
- Law, accounting, insurance, real estate offices. Client intake → fill engagement letters from intake answers, flag conflicts → attorney or CPA reviews and signs; no legal research. Listing prep → pull county records and comparable sales, draft the description → the agent sets the price and checks disclosures.
- Property managers. Maintenance requests → sort by urgency, send to the right vendor, draft tenant updates → manager handles emergencies and approves costs.
- Medical, dental, and vet practices. Reminders and no-shows → read the schedule, send reminders, offer waitlist slots → staff confirm anything medical or financial.
- Offices everywhere (a university office counts if the job is one all offices share). A shared inbox → sort, draft replies from policy with the source cited, route exceptions → a person sends. Event logistics → build the schedule, draft invites, track RSVPs → coordinator approves spending. Forms and PDFs into a spreadsheet → gather, extract, enter, flag anything odd → a person reviews the flags.
Building your assistant, Monday by Monday
None of these is turned in. You build it in class, show a classmate, and write it up in Friday's log. The graded build is version 1 on Oct 19, when all of it has to run your team's job together.
| Wk | You build | You know it works when |
|---|---|---|
| 1 | Setup | It runs on your laptop and you can say what it can see |
| 2 | First skill | It does something a one-off prompt doesn't, and you can say what |
| 3 | First hook | It fires on its own while a classmate watches |
| 4 | Rules | You can show it arguing with you, and the same request without the rule |
| 5 | Memory | Your list of guesses lives in it, and you can say what it forgets |
| 6 | Second hook | The one your project needs, firing on real information |
| 7 | Put it together | All the pieces run your team's whole job on sample information. Version 1 due Mon Oct 19. |
| 8–10 | Helper agents, checker agents, breaking it on purpose | A helper with a job, a checker that catches a failure you planted, twenty inputs passing |
| 11–12 | Run it for real, hand it off | Five days live; a coworker of your contact can run it from your instructions |
By the end of the term: at least three skills each doing one part of your team's job, two hooks (one from Week 3, one your project needs), three rules (one that makes it argue with you), a written memory plan, one helper agent and one checker, and twenty real inputs with a pass/fail check each.