12 January 2026
What to bring to the first class (and what to leave at home)
The most useful object in the first session is a boring piece of your actual week: a reply you send often, a paragraph you always rewrite, a table you dread reconciling. Polish is optional. Specificity is not. If you arrive with a generic “write me a marketing plan,” we will spend the hour turning it into a real task anyway, and you will be behind the people who brought the ugly original.
Leave identifiers at home. A client’s name, a student’s ID, a tenant’s address, a patient’s date of birth, a donor’s story with enough detail to recognize them in a small town — none of that belongs in a classroom chat, even a closed one. We will help you strip the example until it is still useful and no longer anyone’s file. If the task collapses when you remove the person, it is the wrong task for this room.
Bring a note about your constraints. “Our IT has not approved this vendor.” “I cannot install anything.” “French goes to a translator, not to me.” Those sentences save a week of false starts. We teach tool-agnostic method first; the constraint tells us which second layer is even available.
Bring permission, if the work is not solely yours. A manager who expects you to paste the whole drive into a consumer tool is a management problem. We can help you write the sentence that says no. We cannot undo a leak.
Leave the performance at home as well. You do not need to have “kept up.” Foundations exists because many working Canadians have opened a chatbot once, closed it, and felt both late and unimpressed. That is a reasonable place to start.
If you are unsure what is safe, write to us before the session with a one-line description, not the file. We would rather delay your example than let a real person ride along in your laptop bag.
4 February 2026
Where AI still loses to a spreadsheet
A model is at its best when the job is language: a draft, a rephrase, a list of questions you might have forgotten to ask. A spreadsheet is at its best when the job is to be wrong in a visible way. If a total does not add, a person can see it. If a model summarizes a column and quietly drops the awkward rows, the prose still sounds finished.
We keep sending these jobs to a table first: counts, filters, “how many of these are overdue,” “what is the median, not the story.” Asking a chatbot to “look at this export and tell me the trend” is how a rounded number that never existed in the file becomes a sentence in a briefing note. The sentence then gets quoted because it was fluent.
There is a second class of loss: invented structure. Models love to propose categories. Categories feel like analysis. They are often just a sorting hat. If you need groups that your organization already uses — program codes, cost centres, case types — those codes live in the sheet or the system of record. The assistant may help you write a definition. It should not get to invent a new taxonomy on a Thursday afternoon.
When we teach Data Sense Before AI, the turning point is usually a participant watching a confident paragraph fail a simple checksum. After that, people start writing two-step briefs: “do not calculate; wait, I will paste the already-checked totals; now help me explain them in plain language.” That split is the whole course in one habit.
None of this means language models are useless near data. They are useful for naming columns, for drafting a data dictionary, for explaining a pivot you already trust. They lose when they are asked to be the pivot.
If your week is mostly numbers, start with the table, then take Prompt Craft for the paragraph that has to travel with those numbers. Doing it in the other order is how a beautiful memo gets a quiet hole in the bottom.
26 February 2026
Writing an internal AI rule your team will actually follow
Most internal AI policies we are shown are either a vendor’s acceptable-use page pasted onto letterhead, or a legal memo that nobody who answers the phone will read twice. The first kind is not yours. The second kind is not operational. A rule people follow is short, posted near the work, and specific about the files they actually touch.
We start with four buckets in the language the team already uses: public, internal, confidential, and never. “Never” is the useful one. It is not a vibe. It is a list: identifiers, health and student records, unpublished complaints, children’s data, anything still under investigation. If your sector has a statutory overlay, your privacy office writes that overlay. We write the desk version so a person on a Saturday can say no without a seminar.
The second half of a usable rule is about sending, not about drafting. Who may put text into a tool, who approves a letter, who is not allowed to press send because the model sounded kind. Kindness is a risk in complaints. Fluency is a risk in numbers. Spell those out.
Logs should be light enough that people will keep them. Tool, date, class of task, initials. You are not building a surveillance program. You are making it possible to answer a reasonable question next year. If the log is a twelve-field form, it will be empty.
This is practical training, not legal advice. A draft rule from class is a starting artefact. Counsel, a privacy office, or a union agreement may need to mark it. We would rather you leave with a page that gets annotated than with a 40-page PDF that never comes down from the intranet.
If you want a place to begin, sit with the people who do the work and ask: what did you paste last week? Write the rule against that week, not against a future that has not arrived.
19 March 2026
Reading a model's answer like an editor, not a believer
The skill is not prompting. Prompting is the brief. The skill is reading. Models produce text that has the surface of competence: headings, balanced tone, a closing that sounds like it took a position. An editor looks for the sentence that was never in the source, the citation that does not exist, the number that was rounded until it became a different number, the “including” list that quietly dropped the politically awkward item.
In class we print the output. Paper slows people down. We mark claims. We ask “where is this from?” If the answer is “it sounded right,” the line comes out. This feels rude the first time. It is the same rudeness a good editor already applies to a junior colleague, minus the colleague’s feelings.
Style is a second pass, not a first. If you start by making it sound like your organization, you will defend sentences you have not checked. House voice can wait until the facts are nailed to something you hold.
There is a particular Canadian failure mode: confident language about a federal program, a provincial form, or a municipal by-law that changed last year. Models are not your gazette. If the work depends on a current rule, open the rule. The assistant can help you outline a letter once you have the clause in front of you.
We teach a short stack of questions: What would make this wrong? Who gets hurt if it is wrong? Which numbers must be retyped from the source? What must be quoted rather than paraphrased? If you cannot answer those, you are not done, however finished the paragraph looks.
Belief is the hazard because the prose is designed — not with intent, but with training — to be agreeable. Disagreeing with a paragraph that is trying to help you takes practice. That practice is the course.
9 April 2026
Teaching a team that spans three time zones
Canada is wide. A “lunch-and-learn” booked from Toronto is an early-morning sit for Victoria and a late-afternoon sit for St. John’s, unless you pretend otherwise. We run open cohorts on Eastern Time on purpose, with evening and morning streams, so people can choose a slot that does not require them to be the heroic one on the call.
For closed team programs the honest move is to name the time zone in the statement of work and to record with consent. Recording is not a substitute for review. It is how a person in a later zone, or a person covering a counter, still sees the showing. The attempt and the critique still need a live block, even if that block is shorter.
Asynchronous homework has to be designed for fatigue. A task that assumes everyone will wrangle a new login at 21:00 local time will quietly drop the people you most needed in the room — the ones already doing shift work. We prefer homework that can be done in the tools they already open, on a sample they already stripped.
Bilingual teams add a second clock: translation and review. If French or English public text must be checked by a person, do not schedule the “final polish” for the hour after the live session. Machine-fluent copy still needs a human who can hear when the community’s wording has been flattened. We teach in English; we still plan for that extra day.
When three time zones are in one intensive, we cut the heroics. Fewer live hours, more written briefs, a review that is booked twice so no one is always the one eating dinner on camera. If a sponsor wants everyone “together for the energy,” we will say what that costs the people at the edges of the map.
The artefact should not depend on having been in the room for a joke. Notes, templates, and the dated files are the memory. The call is for the parts that need a voice: the failure demo, the line-by-line, the question a person was afraid to type.