The most expensive sentence in an AI workspace is: you know what I mean.
Assistants do not know what you mean. They know what you wrote, plus whatever context they can verify. The craft of delegating to one is the craft of writing the sentence that leaves no room for a confident wrong guess.
Include the outcome, not the steps
Compare two briefs. First: update the pricing tracker. Second: update the pricing tracker so it reflects the March vendor sheet, flag rows where the delta is over ten percent, and leave the notes column untouched.
The second brief constrains the work. When a constraint is impossible, the assistant can say so. When nothing is constrained, every interpretation is technically valid, including the wrong one.
State what must not change
Negative constraints are cheap and powerful. Do not email anyone. Do not delete anything. Do not touch the archive folder. An assistant that knows the fences will stop at them and ask, instead of strolling through.
Ask for questions
End briefs with: if anything is ambiguous, ask before starting. This one line converts silent guessing into a short clarifying exchange, which is almost always faster than discovering the wrong assumption in a finished draft.
Good briefs look like good tickets. Specific outcome, clear boundaries, explicit unknowns. The assistant is not the variable here. The brief is.