The trap: asking AI to build too soon
Most bad AI work starts with a fast answer to a half-understood problem.
We say, "Build me an outreach agent." AI hears the noun, races toward the deliverable, and fills every missing fact with a reasonable-looking guess. Ten minutes later we have folders, workflows, and copy. None of it knows what good outreach looks like in our business.
That isn't speed. It's a contractor pouring concrete before asking where the house goes.
The move is simple: don't ask AI to build yet. Ask it to understand first.
The conversation before the construction
- 1Interview meUnderstand the business result, the people involved, what success looks like, what we've tried, and what could go wrong.
- 2Inspect the evidenceRead the successful examples, the misses, the tools, the accounts, the files, and the history that show what good looks like here.
- 3Show the gapsName what is still unclear instead of quietly guessing.
- 4Recommend with rationaleExplain the plan, the tradeoffs, what was rejected, and why this path deserves the next step.
- 5Earn the next permissionGet approval before creating files, connecting systems, contacting people, or taking action that isn't easy to undo.
Clarifying questions prevent the wrong work. Rationale lets us inspect the thinking before the work becomes real.
The prompt
The interview-first agent prompt
I want your help designing and eventually building an AI agent for [BUSINESS OUTCOME]. Don't build anything yet. First, interview me one question at a time to understand my business, the people involved, what success looks like, what I've already tried, and the risks I care about. Then inspect the examples, tools, accounts, files, and history I can safely give you. Show me the gaps in your understanding. Recommend a plan with your rationale and the tradeoffs. Ask for my approval before you create files, connect systems, contact anyone, or take any action I can't easily undo. Teach me what you're doing as we go.
This isn't a clever incantation. It's a management contract.
Why one question at a time matters
A wall of fifteen questions feels like homework. One sharp question starts a conversation. The answer changes the next question. That's the point.
A good employee doesn't arrive with a questionnaire carved in stone. They listen, adjust, and keep digging until the job is clear.
Give it evidence, not adjectives
"Write good outreach" tells AI almost nothing. "Here are nine messages that earned replies, plus nine that didn't" gives it something it can study.
Ask it to compare them. What patterns show up in the wins? What weakened the misses? What should become a rule, and what was a one-off? Your archive holds your standards in the wild. Let AI read the scar tissue before it invents a playbook.
Rationale is the trust layer
The recommendation matters. The reason underneath it matters more. When AI says what it recommends, why, what it rejected, and what could go wrong, we can manage it. We can correct the assumption before it becomes a campaign.
Trust grows one correct decision at a time. First it explains. Then it proposes. Then it acts with approval. Eventually, after it has earned the right, we stop checking every small move.
Strategy before staffing
Sometimes the interview reveals that we don't need an agent yet. We need to decide what the work should be.
Which buyer matters? What gift would make the first outreach useful? Which successful pattern deserves to scale? What should remain human because the judgment is the advantage? Let AI be a sparring partner before turning it into an employee. Once the strategy is clear, the job description gets easy.
Your homework
Pick one job you've been tempted to automate. Open a fresh AI conversation and paste the prompt above. Don't let it build anything in the first exchange. Answer the interview, give it three real examples of what good looks like, and make it explain its recommended plan. Bring the best question it asked you back to the room.