The unpromptable stack: coaching, judgment, and collaboration skills L&D must protect in an AI-heavy workplace

You can watch someone get better at prompts in a single afternoon. The input gets tighter, the output gets less generic, and the person leaves with a template they will actually reuse. That afternoon is worth running. It is also the smallest part of the skill stack a team will need once the tools are normal.

The rest of the stack does not respond to a better prompt, because it is not a prompt. It is a manager helping someone see their own work. It is a person deciding a fluent answer is still the wrong answer. It is a group turning a pile of options into a decision they will own together. Call that the unpromptable stack. If L&D spends the year on prompt templates and lets this stack thin out, the organisation gets faster drafts and weaker decisions.

Coaching

A model can list feedback in a friendly tone. It cannot watch someone flinch, choose the one point that matters this week, and stay with them while they try again. That is coaching, and it is already scarce in most mid-market teams because managers were promoted for being good at the work, not for teaching it.

The tools make that gap more expensive. When a draft appears in seconds, the temptation is to edit it yourself and move on. The person who brought the task learns that the job is to fetch a draft, not to get better at the work. Six months later the team can summon text, and the manager is still the only person who can tell a good draft from a bad one. You have saved keystrokes and concentrated the judgment in one person.

Protect coaching by giving it a slot that a tool cannot fill. One conversation a fortnight, on real work, with the model output in the document and the human decision circled. The manager's job in that slot is questions. What did you accept too quickly? What would the customer argue with? What will you do differently on the next one? The model can suggest the questions. It should not replace the conversation.

Judgment

Judgment is the refusal to confuse a complete looking answer with a right one. It sits on domain knowledge, on a sense of what the organisation has promised, and on the nerve to slow down. None of that is stored in a prompt library.

The risk in an AI heavy workplace is that judgment atrophies because the first draft is always available. People stop holding the problem in their head. They start holding the prompt. You still need people who can look at a recommendation and know which assumption is doing the work, especially in roles that touch money, safety, customers or the law.

L&D protects judgment by assessing it. Not with a quiz on prompt structure. With a work sample: here is a draft, here is the source, mark what you would not send and say why. Do this for the roles where a miss is expensive. Promote people who are good at it. If the only people who get recognised are the ones who produce the most output, you will train volume and untrain judgment.

Collaboration

Collaboration is not a shared chat thread with a bot in it. It is the messy work of aligning people who see different parts of the problem, including the person who was not impressed by the demo. A model can summarise the thread. It cannot carry the trust, the status and the tradeoff.

Teams that outsource that summary too early start agreeing with a document none of them wrote. The meeting gets shorter and the commitment gets thinner. Protect a part of the work where people still have to think together before a tool tidies the language. A decision meeting with the options on the wall, a disagreement that is allowed to stay visible, a named owner. The tidy version can come after the choice.

What to keep on the curriculum

Write the year as two columns. One column is tool fluency: prompts, approved tools, data rules, a workflow or two. Keep it short and refresh it when the tools change, because they will. The other column is the unpromptable stack, and it should be harder to cut.

1. Managers get a coaching rep on real AI assisted work, not a slide about empathetic leadership.
2. High stakes roles get a judgment sample every quarter, marked by someone who knows the domain.
3. Cross functional work keeps a decision ritual that happens before the summary is generated.
4. Prompt training is offered as a clinic, on demand, rather than as the flagship programme.

The critical thinking workshop belongs in that second column. It is practice at holding an answer up to the problem, which is the piece a template will not do for you.

Let the tools take a first pass on summaries, sorting and empty drafts. Spend the hours you get back on coaching, on judging a live draft and on the conversation a group still has to have before it decides. If those hours are handed straight back to more volume, the old job gets faster and the part that needed a person gets thinner. L&D's job this year is to see which of those two things is happening, and to fund the work that still needs a person in the room.