Claude for Product Managers: Scoring Your Backlog
5 min read
Every backlog review has the same moment. Forty items, a room of people, and a conversation that gets decided by whoever argues best. Scoring frameworks exist to avoid that, and they have their own problem: the score looks objective, the inputs are guesses, and nobody can see which guess moved an item to the top.
Claude can help in a specific and limited way. It can apply a rubric you wrote consistently across many items, show its reasoning for each score, and rerun the ranking under different weights. It cannot know your customers, your strategy or your constraints. This article shows how to use it for the first and not pretend it does the second. For the basics of working with Claude, see the Complete Beginner's Guide to Claude.
The rubric comes from you
If you let an assistant choose the criteria, you will get a polished version of whichever framework it has seen most often. Write your own, even a rough one. Three to five criteria is plenty.
- 1
Choose the criteria that reflect your current strategy
For example: customer impact, strategic fit, effort, risk and how well we understand the problem. Drop any that you would not actually act on.
- 2
Define each one on a 1 to 5 scale, in words
"Customer impact 5: affects most active customers weekly and has a clear link to retention. 1: affects a few users rarely." A defined scale is the difference between consistent scoring and a vibe.
- 3
Set weights, and write down why
Strategic fit 30 percent, customer impact 30, effort 20, risk 10, understanding 10. The "why" tells you later whether the weights still fit.
- 4
Decide what Claude should do with missing information
"If the item does not give enough to score a criterion, mark it unknown and say what is needed." Without this rule, Claude will fill the gap with plausible guesses.
Give it the items in a consistent shape
Scoring works best when each item has the same fields. A short description, who asked for it, the evidence so far and a rough effort estimate from engineering.
Here is my scoring rubric with weights: [paste]. Here are twelve backlog items, each with a description, source and notes: [paste]. Score each item on each criterion from 1 to 5 using only my definitions and the information given. For each score, write one sentence of reasoning that cites the information it is based on. Where the information is not enough, write "unknown" and say what would settle it. Then compute the weighted total and show a table sorted from highest to lowest.
”The "one sentence citing the information" rule makes the scores auditable. When someone asks why item seven beat item four, you can point to a line, not a feeling.
Treat "unknown" as the most useful output
The most valuable column in the result is the list of unknowns. If six of your top ten items have an unknown on customer impact, the ranking is not telling you what to build, it is telling you what to find out.
Tip
Ask for a second table: for each unknown, what is the cheapest way to find out (a customer call, a data query, a prototype)? The research plan is often a more useful result than the ranking.
Stress-test the ranking
A score is only as stable as its weights. Check how much the order depends on them.
Rerun the ranking three times: once with equal weights, once doubling the weight of effort, and once doubling the weight of customer impact. Show how the top five change in each case. Tell me which items stay in the top five in all four versions, and which only appear under one weighting.
”Items that stay near the top in every version are robust. Items that jump around are sensitive to your choices, and they are the ones to discuss.
Look for the bias in the inputs
Scoring is not neutral. Items from the loudest stakeholder or the most recent customer call tend to arrive with better descriptions, and better descriptions score higher.
Look at the information I gave for each item. Which items have detailed evidence and which have only an opinion? Is there any pattern, for example, all the best-documented items come from one team or one customer? Tell me where better-written descriptions might be inflating scores.
”If the answer is yes, fix it before the meeting: ask the quieter teams for the same fields.
What you present, and what you keep
A scored table is a conversation starter, not a verdict.
Present
- The rubric and the weights, so the criteria are visible
- The ranking with the reasoning for the top and the borderline items
- The unknowns and how you would resolve them
- How the ranking changes under different weights
Keep as your judgment
- Dependencies and sequencing between items
- Commitments you have already made
- Anything the numbers cannot capture, such as a risk to a key account
- The final call on what goes into the next cycle
Common mistake
Letting the total be the decision. Ranking forty items by score and cutting at line ten treats a guess as a measurement. Use the score to organize the discussion, then decide.
A habit worth keeping
Save the rubric and the prompt in a Claude project, so each planning cycle starts from the same definitions. Setting Up a Claude Project You'll Actually Keep Using walks through it. Date the weights and revisit them when your strategy changes, because a rubric nobody revises ends up scoring last year's priorities.
If you also gather user feedback before scoring, other assistants offer workflows for it: ChatGPT for Product Managers: From Rough Idea to One-Page PRD and Gemini for Product Managers: Turning Feedback Into a Roadmap Item cover the neighboring steps.
Official sources
Checked on October 24, 2026. Features, plans and names change often, so the vendor's own pages are the final word.