What makes an AI task too large?

Watch the briefing

Watch the full video on YouTube

Briefing

“Organize the exhibition.” What should the answer contain: an inventory, wall labels, or a plan to install the work? Those are different jobs. In our fictional exhibition, the request also hides information we do not have. A useful first move is to choose one deliverable that you can inspect. Start with an inventory. Its job is to organize the supplied facts, not decide the whole exhibition.

Here are our invented notes. Three prints have the identifiers P1, P2 and P3. The first is River Steps. The second is Market Window. The third title is not confirmed. There is no opening date, wall measurement or mounting approval in these notes. An illustrative bad draft says “opening Friday at six” and “hang everything on the north wall.” Those details are not in the input. This mock draft is a teaching example, not a result captured from a named AI tool.

Instead, ask for an inventory table using only the notes. Name the columns: identifier, title and unresolved information. Tell the tool not to invent dates, measurements or approvals, and ask it to return only the table. Our illustrative acceptable version contains the two known titles and a title-to-confirm placeholder for P3. Mounting approval remains unknown. We have narrowed what success means: preserve each identifier, preserve the supplied titles, and keep the gaps visible.

Now inspect this first result against the notes. Are all three prints present? Did the first two titles stay attached to the correct identifiers? Is the third still unresolved? That is a check you can actually perform. If the table changes an identifier, correct that before using it elsewhere. Dividing the task does not make the answer automatically right. It gives you a smaller piece to compare, rather than a polished plan with many hidden assumptions.

Use the reviewed inventory for the next deliverable: label drafts. Ask that each label retain its identifier and exact confirmed title. P3 should stay a placeholder until someone confirms the title; a plausible replacement is not confirmation. The labels depend on the inventory. If a title changes after review, update its label too. The point of splitting the work is not to create more unrelated chats. It is to make the relationship between inputs and outputs clear.

Only then request a preparation checklist. It can identify questions such as who confirms the titles, who approves mounting, and which measurements are still needed. It cannot establish that a wall or fixing is suitable from missing evidence. Keep those decisions with the responsible people and applicable venue procedures. These are administrative teaching examples, not instructions for installing art safely. The prompt boundary and the real-world approval boundary are both necessary; neither replaces the other.

For your next oversized task, name the first artifact, the facts it may use, and the checks it must pass. Review that artifact before it becomes input to the next one. There is no universal word count that makes a task too large; the practical question is whether you can inspect the result and its dependencies. Request one inspectable deliverable at a time. A related lesson asks whether the entire process actually saved time, including the review.