Short pieces I write after a public cohort, dated 2026. They are from the room, from the homework, and from the clinic. They are not a newsletter and they do not sell a seat.
The first week always goes on naming the task
Every January group arrives with a tool already open and a vague wish to “get better at AI”. By Thursday I have learned that the wish is covering a real job that nobody has written down. The report someone mentioned is actually a monthly pack for a director, built from four sheets that do not share a date column. Until we say that out loud, every prompt is theatre.
I now spend the first evening on a boring exercise. Write the task in one sentence that names the input, the output, and the person who will read it. Then write a second sentence on what the model is forbidden to invent. People resist this. It is the only way I know to stop a fluent paragraph that answers a different question.
Homework that week is the same sentence, rewritten after they have looked at their own file. About a third of the room changes the task completely once they see the columns. The later evenings only make sense if the job is small enough to finish. When someone wants to “do all of operations”, I send them back to one loop.
How I check a model’s numbers
On the data course I have a rule I say out loud until people roll their eyes. If the assistant returns a total, you find the rows. If it returns a percentage, you find both sides of the fraction. If it “helpfully” fixes a date, you look at the original cell. Fluency is not evidence. I have watched a model rename January as June because the rest of the paragraph was about mid-year planning.
The check I teach is short on purpose, because a long check will not survive a Thursday afternoon. Highlight the claim. Go to the sheet. Filter. Count. If you cannot reconstruct the number in two minutes, the claim does not leave the desk. People want a more sophisticated method. The sophisticated method is still “go and look”, plus a note in the file that says which range you used.
I also ask them to trick the model once. Give it a sheet with a blank in a column that looks complete, or a duplicate invoice. Watch what it does. The point is not to catch the tool out for sport. The point is to feel, in the body, that a confident sentence can sit on top of a hole. After that, the checklist is less of an argument and more of a habit.
Three places an automation falls over after a month
The Saturday course produces a lot of working loops by 18:00. A month later, in clinic, three failures show up so often I now teach them before lunch. The first is an empty field. The form allowed a blank “customer name”, the later step tried to name a folder, and the run stopped with a message nobody read. The fix is ugly and reliable: refuse to continue, and send the human a line that says which field was empty.
The second is a duplicate. Someone submits twice. The sheet grows a twin row. The automation files two folders and sends two mails. If the loop cannot look for a unique id, it should at least tag the row as “possible duplicate” and wait. The third is permission. A colleague leaves, a token expires, a shared inbox is renamed. The loop fails silently because nobody owns the login. I now make the group write a one-page runbook with a named owner and a date to check the connection.
I will not let a group build five automations in a day. One loop that still runs in March is the whole Saturday.
What to do with confidential files
The first instinct in every group is to paste the whole pack. Contracts, CVs, a customer export with identity numbers still in column F. I do not moralise about this. The tools make it easy, and the week is already too full. I do stop the paste and we do the slower thing: copy the file, strip the names, keep the shape, and only then ask the model to help with structure or tone.
Shape is the useful part. A twelve-page agreement with the parties replaced by A and B will still teach you whether a clause is missing. A sheet with invented amounts will still teach you whether a variance flag is firing on the right rows. What you cannot reconstruct later is trust, once a real person’s data has gone into a public model your organisation does not control.
Some jobs should stay offline. Anything that identifies a child, a patient, or a bank account belongs in that pile until your own security people have chosen a private tool. On the half-day course we write a one-page rule in the language the team already uses, and we practise saying no to a cheerful request to “just drop the drive in”.
There is no perfect prompt
People arrive with a rumour that somewhere on the internet there is a paragraph which, once pasted, makes the model behave. A prompt is a brief. Briefs go stale when the audience changes, when the source file grows a new column, when the writer who owned the voice goes on leave. Treating one block of text as a finished instrument is how you get polite, empty copy that still needs an hour of editing.
What I teach instead is a short loop. State the job. Give one example of the output you like. Name two things the model must not do. Generate. Check a fact. Tighten one constraint. Keep the draft in a document you can edit. A chat thread disappears. The “prompt pack” you leave with is a set of those loops, written in your words, for your genres. It will look modest. That is the point.
If a trick helps, we keep it. If a trick only works on a demo, we throw it away in front of the room so nobody feels they missed a secret. The clinic questions in March have all been some version of “why did this stop working”. The answer is usually that the world moved and the brief did not. Update the brief. Do not hunt for a longer spell.
Finance asks different questions from marketing
When I mix functions in a public group I can feel the room split at the first homework. The analysts want the model to show its working. They will forgive clumsy prose if the rows are named. The marketers want the model to stay in voice. They will forgive a soft number if the paragraph still sounds like the brand. Both are reasonable. Homework stays on each person’s own files.
Finance questions I hear every cohort: Can I trust this total. What did you do with the blank cells. Who signs the pack if a figure is wrong. Marketing questions: Can it sound like us without copying a competitor. How do I stop it inventing a customer result. Where does a testimonial come from if nobody said it. The checklists overlap on one point only. Do not let a fluent sentence leave the desk until a human has decided it is allowed to exist.
I like teaching both in the same room because they hear each other. An analyst listening to a brand argument learns why “just make it nicer” is a real constraint. A writer listening to a variance argument learns why a rounded percentage can be a small lie. Then we go back to our own laptops.
I use a single local storage key to remember this choice and anonymous traffic counts. Nothing is sold or shared. See the cookie note.