Teaching a Team to Use AI Well in Two Hours

Two hours is enough to change how a team works with this, provided the session is built on their own tasks and produces a written rule they agreed to. A demonstration of impressive capabilities is not, and it is what most sessions are.

What a session like this is for

Three outcomes, and it is worth stating them at the top of the session so people know what they are there for.

• Everyone has done one real task from their own job and seen it work or fail. Not a demonstration on somebody else’s example.
• Everyone has seen it be confidently wrong, in a situation where it did not matter. This is the single most valuable two minutes of the session and it has to be experienced rather than asserted.
• The team leaves with a written rule about what may and may not go into these tools, that they wrote.

What it is not for: teaching prompt syntax, comparing products, or predicting the future of work. The organisational picture around the session — who runs it, how it fits an adoption programme — is in training your team to use AI well.

Preparation: two hours, the week before

1. Ask each attendee for one task they do repeatedly that involves reading or writing, and roughly how long it takes. One sentence each, by email. This gives you the material for exercise one and it is the reason the session will land.
2. Find out what the organisation actually permits, in writing, before you stand up. If there is no policy, say so in the session rather than improvising one; the rulebook exercise then becomes the first draft of it.
3. Prepare one deliberate failure. Something in your own field where a confident answer is wrong — a fabricated citation, an invented internal policy, arithmetic over a long list. Test it beforehand, and have a fallback, because the failure you planned sometimes does not reproduce.
4. Sort out access. Everyone logged in, on a permitted tool, before the session starts. Twenty minutes of a two-hour session lost to accounts is the commonest way these fail.

The timed outline

0:00  What this is and is not                          10 min
      One paragraph on what the technology does: it predicts
      likely continuations of text. No history, no hype, no
      predictions. State the three outcomes for the session.

0:10  Exercise 1 - your own task, badly                  20 min
      Everyone does their own task with a one-line prompt.
      Then they read out what they got. Expect disappointment;
      that is the point of doing it first.

0:30  The five moves                                     15 min
      Audience, material, shape, example, constraints. One
      before-and-after on a task from the room.

0:45  Exercise 1 again, properly                         20 min
      The same task with the five moves. Compare. Two or three
      people read both versions out.

1:05  Break                                              10 min

1:15  Exercise 2 - make it be wrong                      20 min
      Your prepared failure, then everyone tries to produce a
      confident wrong answer in their own domain. Collect them
      on a whiteboard. This section is the reason the session
      exists.

1:35  Exercise 3 - what we will not put in                15 min
      Small groups. Three things from our work that must never
      be pasted in, and why. Groups report; you write the list.

1:50  The rulebook and what happens next                 10 min
      Agree the rules on the screen. Name who owns the document
      and when it is reviewed. One volunteer per team to report
      back in a fortnight.

The structure has one deliberate feature: the first exercise is designed to go badly. People who arrive sceptical are vindicated, which buys their attention for the second attempt, and people who arrive over-enthusiastic get a useful correction early.

The three exercises

Exercise 1: your own task, twice

First pass: one line, no context, as everybody actually does it. Second pass, after the teaching: the same task with the five moves — audience and next step, the actual material pasted in, the shape of the answer, one example, and a list of what it must not do. The full method is in how to write a prompt that gets a useful answer; you are teaching it in fifteen minutes with one worked example.

The comparison does the teaching. Do not explain why the second is better; ask the person who wrote it.

Exercise 2: make it be wrong

Ask everyone to get a confident wrong answer about something they know well. Reliable prompts for this: ask for the source of a specific claim, then check the source exists. Ask about a niche internal procedure. Ask for a total over a long list of numbers. Ask about something that happened last month.

Then, and this is the part to leave time for, ask how you would have spotted it if you had not known. The answer is usually “I would not have”, which motivates the checking routine. Give them the five-minute fact-checking routine as the takeaway and walk through it once on one of the room’s own examples.

Exercise 3: what we will not put in

Groups of three, ten minutes. “List three things from our work that must never go into one of these tools, and say what the harm would be.” The lists come back specific and correct: client files, candidate CVs, the unreleased figures, the safeguarding notes.

A rule people generated is followed. A rule circulated from elsewhere is not, and this exercise takes ten minutes to produce something a policy working group takes a month to write worse.

The house rulebook

One page. Written on the screen in the last ten minutes, from what the groups produced. This is a starting template; the value is in the amendments the room makes to it.

HOUSE RULES - AI TOOLS
Agreed [date] by [team]. Review [date + 6 months]. Owner: [name].

NEVER PUT IN
  1. Anything identifying a client, a patient or a candidate.
  2. Unpublished financial figures.
  3. Anything from a file marked confidential.
  4. Colleagues' personal information, including anything told to
     you in confidence.
  5. [the team's own additions]

ALWAYS
  6. Check any fact, name, figure or citation before it leaves
     this team. The checker is the sender, not the reader.
  7. Say when a document was substantially drafted this way, to
     colleagues. We do not tell customers unless asked, but we do
     not deny it either.
  8. Use the [approved tool] account, not personal accounts.

NEVER USE IT TO
  9. Make a decision about a person: hiring, discipline, credit,
     eligibility. Draft the letter; do not choose the outcome.
 10. Produce anything going to a regulator without a named human
     reviewer.

IF SOMETHING GOES WRONG
 11. Tell [name] the same day. Nobody is disciplined for reporting
     it. This rule is what makes the others knowable.

THINGS WE ARE NOT SURE ABOUT
 12. [keep this section; it is the honest one]

Rule 11 is the one to defend if it is challenged. A team that hides mistakes made this way learns nothing and the organisation finds out from a customer. Rule 12 keeps the document credible: a rulebook with no open questions was not written by anybody who had used the tools.

Adapting it to the room you have

The outline assumes a mixed room of willing adults in one place. Four common variations need a change, and each change is small.

One thing does not vary: everybody does exercise 1 on their own task, on their own machine. A session where people watch somebody else type produces agreement in the room and no change in the work.

What to leave out

• How the technology works. Two sentences is the correct amount. Nobody has ever written a better prompt for knowing what a transformer is.
• Product comparisons. They date instantly, and the organisation has probably already chosen. Teach the method; the tool is an implementation detail.
• Predictions about jobs. Either reassurance or warning, both are unfounded and both cost you the room. If somebody raises it, say honestly that nobody knows and return to the task.
• Impressive demonstrations. The party trick is memorable and teaches nothing. Everyone has seen one.
• A slide deck. Two hours, three exercises, one document written live. If you are talking for more than thirty-five minutes in total, the design is wrong.

The week after

Sessions decay unless something structural changes. Three cheap things hold the gain.

1. Circulate the rulebook within twenty-four hours, while people remember agreeing to it, with the names of who contributed which rule.
2. Start a shared file of prompts that worked, with a one-line note on what each is for. This is the artefact that makes the second month better than the first, and it costs nothing.
3. Fifteen minutes in a team meeting a fortnight later. What worked, what wasted time, what surprised you, and what should change in the rules. The second conversation is where the honest assessment appears, because by then people have tried it on real work and stopped being polite about it.