Latent-expertise prompt-pattern guide
AI amplifies judgment you already have — it doesn't manufacture judgment you don't. This guide walks through the chapter's method for picking one expert, one task, and building the prompt pattern that lets their expertise move faster without leaving the loop.
The three skills that determine whether this works
None of them are technical. All three are management skills your best operators have been building for years.
- Explaining what you actually need — a precise specification of the task, the constraints, and the standard the output must meet. Not a vague gesture at an outcome.
- Evaluating the output — the model will hand you something confident and plausible every time. Whether it's right is a question only someone with domain knowledge can answer.
- Giving feedback — taking the 70%-right output and steering it to 95% through a couple of rounds of correction. The same muscle every good manager uses with a capable junior employee.
Step 1 — Pick the expert and the task
Don't roll AI out to everyone at once. Pick one person with deep knowledge of one specific high-value workflow, and the one task where their judgment is the bottleneck.
| Selection question | Your answer |
|---|---|
| Who on the team has the deepest knowledge of a specific high-value workflow? | |
| What is the one task where their judgment is the bottleneck — the analysis only they can do well, the review only they can be trusted with? | |
| What is the current output rate for that task (e.g., 2 reviews per day)? | |
| What would "10× faster without getting worse" look like? |
Step 2 — Map the latent expertise
Before writing a single prompt, get the expert to articulate the judgment they apply. This is the step most teams skip, and skipping it is why the output looks polished but isn't right.
| Question | Expert's answer (in their own words) |
|---|---|
| What is the task, exactly, in one sentence? | |
| What does "good output" look like? What's the standard you're judging against? | |
| What are the two or three most common ways a naive person would get this wrong? | |
| What context does the model need to reason the way you reason? | |
| What are the edge cases the model must flag rather than answer? |
Step 3 — Build the prompt pattern
Fill in the template below. This is the working artifact: the expert reviews each line against what they actually know, not what sounds good.
Expert: ______
Task: ______
The judgment they apply (in their words from Step 2):
______
The standard the output must meet:
______
The most common errors to guard against (list them explicitly in the prompt):
______
The edge cases the model must surface rather than resolve:
______
Draft prompt pattern (write it here, then test it on a real example):
______
Step 4 — Iterate until the expert would sign their name to it
- ☐ Run the prompt on three real examples from recent work.
- ☐ Expert evaluates each output: what's right, what's wrong, what's missing?
- ☐ Revise the prompt based on the gaps. Repeat until the output is something the expert would send without editing.
- ☐ Record the final prompt pattern. It's now a firm asset, not a personal trick.
The trap to avoid
AI does not lower the bar for who can do expert work. It raises the ceiling for people who already clear the bar. Do not hand this prompt to someone who doesn't have the underlying expertise and assume the tool supplies it. The only thing standing between "plausible" and "right" is a human who can tell the difference.
Want a second set of eyes on this in your firm? The no-sell promise applies — if it isn't a fit, I'll tell you in the first ten minutes.
Book a 30-Minute Call →