Pre-launch · Training and the free Foundations exam are open now. Levels 1–5 open soon. Get notified on any level page.
20 min read
A rubric is a short list of what "good" means for a particular output, written down before you read it. Without one, review becomes a reaction: it feels right, so it ships. With one, you and a colleague can review the same draft and reach the same verdict. This module shows you how to build a simple rubric, use it to compare options fairly, and check the numbers behind a recommendation.
A simple rubric has three to five dimensions. Each dimension states what you're checking and what passing looks like. For a customer reply, that might be:
Write it before you read the output. If you write it after, it tends to describe whatever you got.
You can give the rubric to the model and ask it to check its own draft. That's useful for catching obvious misses. It's not a substitute for your review, because the model may grade itself generously or miss the same error twice.
Comparing options. When the task is choosing between alternatives, the common failure is describing each option on different terms: one on price, another on features, a third on reputation. That makes the choice look easy without making it fair. Decide the criteria first, then require every option to be rated on every criterion, with "not supplied" where data is missing. Mark which criteria are must-haves. An option that fails a must-have is out, no matter how well it scores elsewhere.
Evidence and assumptions. A recommendation should show its work: which facts it relied on, where they came from, and what it assumed. "Vendor B is cheapest" should come with the numbers and the source. "Assuming 40 trips per year" should be labeled as an assumption, so you can change it.
Numbers. If a recommendation depends on a calculation, ask for the method: each quantity with its unit, the formula, the result, and a rough estimate to compare against. Then check it yourself with a calculator or spreadsheet. Unit mix-ups such as monthly vs. annual, per person vs. per group, or dollars vs. thousands of dollars are common and easy to miss in fluent prose.
Grace Lindqvist coordinates travel for a 60-person engineering firm. She's choosing among three travel booking services and has quotes in hand. The firm expects about 45 trips a year. Must-haves: 24-hour phone support and an expense-system export.
Weak approach. Grace pastes the three quotes and asks, "Which travel service should we pick? Explain why." The answer praises Service A's interface, notes that Service C is "budget-friendly," and recommends A. It never mentions phone support. It quotes Service B's fee as $12 without noticing that B charges per traveler per trip, while A and C charge per trip.
Better approach. Grace writes the rubric first and makes the comparison and the math explicit.
Before:
Which travel service should we pick? Explain why.
After:
Compare Services A, B, and C using only the attached quotes.
For every service, report each criterion below. Write "not supplied"
if the quote doesn't say.
1. 24-hour phone support (must-have): yes / no / not supplied
2. Expense-system export (must-have): yes / no / not supplied
3. Booking fee: amount, currency, and what it's charged per
4. Annual fee cost: show the formula using 45 trips per year and
an average of 1.5 travelers per trip. Label both as assumptions.
5. Cancellation terms, quoted from the source
Then: list services that pass both must-haves. Recommend one only
from that list, and cite the quote page for each fact you rely on.
Why this is stronger. Every service is judged on the same five criteria. Must-haves filter before preferences get a say. The fee calculation shows its units and assumptions, so Grace can see that B's $12 per traveler becomes $12 × 1.5 × 45 = $810 a year, not $540. She checks that figure in a spreadsheet and confirms each must-have against the quote page cited before forwarding the recommendation.
Setup. Choose a small decision with two or three options: two software trials, three caterers, two shipping carriers. Gather what you know about each in a text file.
Task. Before asking any AI tool, write a four-criterion rubric and mark one criterion as a must-have. Include at least one criterion that needs a calculation with units. Then ask a chat tool to compare the options on exactly those criteria, show its formula, and label assumptions.
What a good result looks like. Every option has a value or "not supplied" for every criterion. The calculation shows quantities, units, and the formula, and your own spreadsheet check matches. Any option that fails the must-have is excluded from the recommendation.
A rubric that grows to 15 dimensions stops being used. If you can't apply it in two minutes, cut it.