AI’s knowledge is like an ocean. When you ask it a broad question with no direction, you’re essentially asking it to boil that ocean. You’ll get an answer back, it will probably sound smart, and it will rarely be exactly what you need.
The people I see getting real value from AI do something different: they focus its intelligence like a missile. The more specific the detail, the better the output. The more expertise you point at one well-defined task, the better the result.
That’s the thinking behind the four-step method I use every time I work with AI, whether I’m building an app, drafting a strategy or analyzing a dataset:
- Persona — Who is the AI?
- Context — What does it need to know to execute well?
- Detailed Prompt — What, exactly, are you asking for?
- Validation — Is the output actually right?
To make each step concrete, let’s use one running example. Imagine you’re an academic dean at a community college building next spring’s schedule, and you want AI’s help deciding which course sections to add, merge or cut.
Step 1: Persona
Start by telling the AI who to be. A persona sets the expertise, vocabulary and judgment the AI brings to the task. Skip it, and you get a generalist’s answer to a specialist’s problem.
Weak: “Help me with my spring course schedule.”
Strong: “Act as an experienced academic operations analyst who has helped community colleges build course schedules that balance student demand, faculty capacity and budget. Your goal is to help me make data-informed decisions about which sections to add, merge or cut for spring.”
Step 2: Context
This is where most people under-invest, and it’s where the heavy lifting happens. Give the AI everything a sharp new hire would need to do this job well on day one: how your organization works, the terms you use and what they mean, how your team and the work are structured, and what success looks like.
Weak: “We have a lot of under-enrolled sections.”
Strong: “We’re a public community college, and most of our students attend part-time. ‘Under-enrolled’ means fewer than 12 students at census. Full-time faculty teach five sections per term, and adjuncts are capped at three. Last spring we cancelled a number of sections in the final week before classes, which disrupted students who had already planned their schedules around them. Attached are three years of section-level enrollment, fill rates by time slot and retention rates by program.”
For a complex analysis, my context alone often runs ten pages or more. The more the AI understands about your world, the less it has to guess.
Step 3: Detailed Prompt
Now make the ask, and make it precise. Define “good” in measurable terms, describe the format you want back, and show examples of what a strong answer looks like.
Weak: “What should we do about our schedule?”
Strong: “Recommend which sections to add, merge or cut for spring. ‘Optimal’ means keeping projected fill rates at 75% or higher while protecting every course required for on-time completion in our five largest programs. Return a table with the course, current sections, recommended sections, projected fill rate and the rationale for each change. Show your reasoning for every recommendation, and flag any recommendation where the data is thin.”
Step 4: Validation
This is the most important step, and the one most often skipped. AI is excellent at producing smart-sounding, legit-looking output. Your expertise is what tells you whether it’s right. Without that check, you can confidently act on something that’s subpar at best and wholly inaccurate at worst.
Weak: Skimming the table, deciding it looks reasonable, and pasting it straight into the memo to your provost.
Strong: Interrogating the output before you trust it: “What assumptions did you make? What could be wrong with the data you used? Which recommendations are you least confident in, and why? What would change your answer?” Then checking the key numbers against your source reports yourself. For high-stakes decisions, go one step further: give a second AI the same data and context, and ask it to poke holes in the first one’s conclusions.
What this looks like in practice
Here’s how I applied the method to a recent analysis of more than 100,000 records:
- Persona: I asked Claude to act as a business analyst and data researcher whose goal was to help me find the strategy that would maximize outcomes.
- Context: I provided at least ten pages of background — how the organization works, the specific terminology and definitions, how the team and the work are structured, and the metrics we expect.
- Detailed prompt: I asked for the optimal strategy to maximize engagement for a specific population, defined “optimal” as 10% or more year-over-year growth in engagement based on specific reports, and included examples of the tables and conclusions I wanted, along with what good and not-so-good looked like.
- Validation: Once I pressed enter, the real work began. I asked what assumptions it had made, what could be wrong with the data it pulled and where the holes were. Then I gave a second AI agent the same data and had it cross-reference the same reports to challenge the first set of conclusions.
Validation is the most important part of the whole process, and it’s what turns an interesting analysis into one I’m willing to put my name behind.
The bottom line
AI multiplies whatever you bring to it. When you bring a clear persona, rich context, a precise ask and a rigorous review, you get work you can stand behind.
This is the method I teach in my AI Enablement Workshops, where every participant applies all four steps to their own work during the session. If your team is ready to get more out of AI, I’d love to talk.
