AI at Work
Draft for instructors — not yet scheduled
Five one-day, no-code classes, shared here for instructors to read before anything is scheduled. Each class page holds the whole class: syllabus, the six blocks, handouts and videos. Each atlas is the same class drawn as one zoomable map.
From Meetings to Action
The Commitment Tracker
Your meetings end. Your action items shouldn't disappear.
AI for Spreadsheets & Reporting
The Monday Report
Stop spending Friday building the report you needed Monday.
Build Your AI Research Department
The Research Desk
Stop asking AI random questions. Give it a research assignment.
Build an AI Outreach & Follow-Up System
The Pipeline
Writing the perfect email is useless if you don't know who to send it to—or what happens next.
Turn Documents into Working Systems
The Policy Desk
Your organization has 400 pages of policies. Can anyone actually find the rule they need?
Build Your Personal Life System
The Life System
Nobody can sell you a life OS. You have to build your own.
Series overview
Six one-day classes. In five of them you build a machine that does a piece of your actual job, on your actual material. In the sixth, the machine runs your life. Every time, you run it twice, then build a second one without help.
No code. Bring your real meeting, spreadsheet, research question, contact list or document set. If you can't bring yours, every class has a practice pack of real public material: a city council transcript, a city's open data, a union contract. Never a toy.
| Class | You leave with |
|---|---|
| From Meetings to Action | The Commitment Tracker. Meeting → owned, dated, sourced action items → your board → follow-ups → the Friday report |
| AI for Spreadsheets & Reporting | The Monday Report. Raw export → a cleaning recipe that re-runs → checks → anomalies explained → charts → a one-page decision memo |
| Build Your AI Research Department | The Research Desk. Assignment → source ledger → citation audit → contradictions → critics → a cited brief → a watch list |
| Build an AI Outreach & Follow-Up System | The Pipeline. Who → a sourced contact list → segments → messages you approve and send → CRM → follow-ups → an honest report |
| Turn Documents into Working Systems | The Policy Desk. Documents → a test set → cited answers that say "I don't know, ask X" → obligations → version diffs → exceptions routed to a human |
| Build Your Personal Life System | The Life System. What it's for → where your passwords, money and health data are allowed to go → your values held up against last week → a plan that fits your real hours and bends for flares and good reasons → its boundaries and yours → a weekly review of what you said vs what you did. A one-day setup, then six student-led LifeOS meetups |
Ask → Assign → Run
Most people stop at the first rung. You type a question, you get an answer, and next week you type it again. Organisations with engineers skip straight to the third rung: in Anthropic's Economic Index (as of September 2025), 77% of enterprise API use of Claude was automation. This series puts the third rung in your hands without an engineer.
- Ask: one chat.
- Assign: a reusable instruction set (your Project, with a Spec Card at the top) and a check you run yourself.
- Run: the same machine on a cadence, with a human who signs before anything leaves.
Every machine is the same seven parts
INTAKE → STRUCTURE → CHECK → GATE → ACT → LEDGER → RHYTHM
That's why the last block of every class works. Once you've built one, the second is a new spec card, not a new skill. Start every class with the Spec Card.
How we assess
You show the machine working. There's no quiz. Every block ends with a check you run yourself: a hand count, a planted error, a citation audit, a test set. Your portfolio is one link to the machine, and the facilitator looks for five things: the spec card in its instructions · two real runs · one failure its check caught · the human gate named · a second machine for a different case.
These replace Using Large Language Models
The foundations from Using LLMs still run through every class: how models see text, picking a model, custom instructions and memory, the flipped interaction, naming the criterion instead of saying "better," expert personas. You just learn them on a real job instead of a tour. Take any class first.
Using LLMs' personal material went to Build Your Personal Life System: the values card sort, affective use, health and clinical language, privacy and memory, and the daily practice protocol. It's also the easiest first class, because you're the only stakeholder in your own spec card. It's the one class that doesn't end at 4pm: six student-led LifeOS meetups follow it, and it runs once a quarter.
The Spec Card
Every machine in AI at Work starts as one of these, written before you type a single prompt. You write it again at the end of the day for your second machine, and the second one comes easier. That's the point.
A spec card is the description of a piece of your job that repeats. If you can't fill a slot, you've found the reason this work keeps going wrong, with AI or without it.
| Slot | The question | Example: the Commitment Tracker |
|---|---|---|
| 1. Intake | What raw thing arrives, and how often? | The transcript of the Tuesday ops meeting, weekly |
| 2. Structure | What record must come out? Name the fields before you look at the source. | Decision · owner · deadline · source quote · status |
| 3. Check | How will you know it's right without trusting it? | Ten minutes of hand-marking the transcript, then count misses and inventions |
| 4. Gate | Which human says yes before anything leaves? | Each owner confirms their own items |
| 5. Act | Where does the result go? Where does the work already live? | The team's Asana board |
| 6. Ledger | What record proves what happened? | The board's history, plus the decision log |
| 7. Rhythm | When does it run again, and who notices if it doesn't? | Tuesday 2pm, a scheduled task; the Friday report flags a missed week |
Rules
- Fields before reading. If you let the model choose the fields after seeing the source, it'll describe the source back to you. That's a summary, not a record.
- The check can't be "it looks right." "Good" isn't a quality. It's a sound you make with your face. Name the number you'll count.
- The gate is a person with a name. "Someone reviews it" is how nobody reviews it.
- Act where the work already lives. Don't create a new place you have to remember to look.
Put it inside the machine
Paste the finished card at the top of your Project's (or custom GPT's) instructions under the heading
SPEC CARD. The machine now knows its own job, and anyone you share it with can read what it's for.
SPEC CARD
Intake: ...
Structure: ... (fields fixed; never add or rename a field)
Check: ... (tell me which items you are least sure of and why)
Gate: ... (draft only; nothing is sent, posted or changed until <name> says yes)
Act: ...
Ledger: ...
Rhythm: ...
Videos: Fundamentals (every AI at Work class, Block 1)
The pieces of Using Large Language Models that every class reuses. Times are ±30 s, so trim by eye before you publish a chapter. Full accounting of every ULM recording: Video Map - ULM Retirement.
| Idea | Clip | Where |
|---|---|---|
| Why it can't count the r's in "strawberry" (tokens) | Exploring Values, Creativity, and AI | https://www.loom.com/share/4faee411dbb744af8b7e09f08f37584d 0:15:20–0:21:00. Best single clip; ends on "it can write Python to count the r's" |
| Same, the ULM lecture | Intro to Using LLMs (ULM W1, 2025-08-12) | https://www.loom.com/share/3cab01a031c442c29233fe2911d70fd3 0:05:50–0:18:00 |
| Picking a model | Choosing the Right AI Tools for Your Business Success | https://www.loom.com/share/326e5d20ebd949b4bd536ed5d25974f2 0:00:00–0:13:30 |
| Same prompt, two models, side by side | Understanding Chat GPT for Prompting and Reasoning | https://www.loom.com/share/e9ce7fbd1e8f40cf831810c339c4e10f 0:30:20–0:40:00 |
| Custom instructions | Setting Up Your AI Assistant: A Step-by-Step Guide | https://www.loom.com/share/feeef3c20bf0448ab34b36bc2b209460 0:03:35–0:17:45 |
| Memory: on, reviewed, pruned (ChatGPT and Claude) | Memory setup (2026) | https://www.loom.com/share/0da504ca33f340d8abd0f0d2bd5e7a51 whole video, 9 min |
| The flipped interaction ("ask me five questions") | Same | https://www.loom.com/share/0da504ca33f340d8abd0f0d2bd5e7a51 0:05:45–0:08:30. Longer: https://www.loom.com/share/feeef3c20bf0448ab34b36bc2b209460 0:15:40–0:28:30 |
| "Better" means nothing: name the criterion | LLM in Analysis and Critique | https://www.loom.com/share/7a69024b78b942698786dce91031c1c6 0:07:30–0:16:50 |
| Projects hold context; long chats rot | Creative Automation session | https://www.loom.com/share/89e95613e5b94c1abf8a3af8bfb28297 1:07:15–1:16:50 |
Ageing: concept segments (tokens, vocabulary, personas) age well. Settings walkthroughs don't, so only the two 2026 recordings (0da504ca, feeef3c2) show current settings screens. Cohort recordings have students on mic, so check consent before publishing outside the cohort.