Anyone who teaches AI for a living learns the same lesson from their own slide decks: the screenshots die first. The interface in the picture gets redesigned. The model in the title gets replaced. The feature being demonstrated becomes a button with a different name in a different menu. Materials that were accurate in March need surgery by June.
This usually gets treated as a logistics problem for trainers. It’s bigger than that. If the content of a workshop can expire in a quarter, then most of what fills AI workshops (which tool to click, in which order, to produce which effect) was never worth a team’s morning in the first place.
If what you taught expires with the next model release, you taught the wrong layer.
The layer that expires
Hold a typical AI-training agenda against six months of release notes and the problem is plain. Model names rotate faster than procurement cycles. Context limits double. Pricing moves. The tool that was leading in autumn has been overtaken by spring. A workshop built from this layer is a tour, and tours are pleasant and forgettable.
The tell is the agenda itself: when it reads as a list of products, the room is being taught things with a shelf life of weeks. The other tell is the demo that only works on the instructor’s examples. Impressive prompts, polished results, no transfer. The participants watched someone else be good at it.
That doesn’t mean a workshop should avoid real tools. You can’t teach briefing in the abstract, and nobody learns checking by discussing it. The tools have to be in the room. The question is whether they’re the message or the medium.
The layer that lasts
What stays true across releases isn’t a feature set. It’s a short list of capabilities that live in the person:
Briefing, not asking. The biggest gap between weak and strong users of these systems is what they provide before the question. Models answer from what they’re given. The skill is assembling context the way you’d brief a capable new colleague: the background, the constraints, what good looks like, what to avoid. That was as true of GPT-4 as it is of whatever shipped this morning.
Checking, not trusting. Outputs are drafts. The durable skill is knowing what failure looks like (confident, fluent, wrong) and where to look for it: the citation that doesn’t exist, the number that isn’t in the source, the step a careful colleague would have questioned. People who learn this stop being afraid of the tool and stop being burned by it, both at once.
Decomposition. Big tasks fail, and bounded steps with checkable outputs succeed. That’s prompt-craft at the small scale and system design at the large scale. Someone who can split a messy job into steps a person can verify has learned something no release invalidates.
Calibrated trust. What to hand the model today, what to keep, and the habit of re-testing that boundary, because it moves. Any verdict about what these systems can’t do comes with an expiry date. Teach the re-test, not the list.
What this changes in the room
Run the session on the participants’ own work, with whatever tools are current, and be explicit about which layer is which. When we show a specific product, we say out loud that this part has a sell-by date and the point being made with it doesn’t. It tends to be the most trust-building sentence of the day.
It changes the take-home, too. A feature sheet ages like the screenshots. What’s worth keeping is patterns: the briefing structures that worked on this team’s tasks, the checking habits matched to this team’s risks, and a short list of what to try next with a date on it for re-asking the tooling question. Recommendations decay. It’s why our own recommendations post carries a date.
And it changes the buying question. Every workshop ends with some version of “so which tool should we get?” The honest answer comes with a date attached: what we’d pick this month, and how you’ll know when that answer has gone stale. A trainer who answers without the date is selling certainty nobody has.
The tools will keep changing every week. That isn’t why training fails; it’s why training aimed at the durable layer is worth a room’s time at all. If a model release tomorrow would invalidate the workshop, it was teaching the wrong things.