How to refine a messy backlog before sprint planning (without staying late)

It's Thursday afternoon. Sprint planning is Monday morning. You open your Azure DevOps backlog and find what every product owner finds: a column of titles that made sense to whoever typed them, and almost nobody else.

"Fix login bug." "Improve onboarding." "Customer wants reports."

You know the drill. Open each one, figure out what was actually meant, write a real description, add acceptance criteria, maybe break it into child tasks, repeat fifty times. By the time you finish, you've spent your entire Friday writing instead of thinking — and half the items are still ambiguous because you ran out of time to chase down the original requester.

This post is a how-to for getting through that refinement pass quickly using ClearSpecs AI, an Azure DevOps extension built specifically for this workflow. The goal isn't to have AI write your backlog for you. The goal is to stop spending your Friday on the mechanical part of refinement so you can spend it on the parts that actually need a human.

Step 1: Set up prompt templates once, use them forever

The biggest time sink in refinement isn't the writing — it's deciding how to write each time. What sections does a user story need? How detailed should acceptance criteria be? Does the team prefer Gherkin or bullet points?

Prompt templates solve this. You define them once in the Boards area, and they appear on every work item form for everyone in the organization. Each template includes the instructions to the AI plus the target field it writes to, so a "Fill description" template knows to write to the Description field, and a "Generate acceptance criteria" template knows to write to the Acceptance Criteria field.

A few templates worth setting up before you start refining:

A "Fill description" template that takes the title and any existing context, then produces a short user story in your team's preferred format (who, what, why). Specify your format in the prompt. If your team writes "As a [role], I want [capability] so that [outcome]," put that structure in the template instructions and you'll get it back consistently.

A "Generate acceptance criteria" template that produces 3–7 testable bullets. Tell it explicitly that each bullet must be something a tester can verify with a yes or no. This single instruction eliminates most of the vague "should work well" criteria you'd otherwise have to rewrite.

A "Definition of ready check" template that reviews the work item and lists what's missing — unclear scope, no acceptance criteria, no linked designs, no mention of edge cases. This one's a refinement accelerator: instead of reading each item carefully looking for gaps, you get a checklist.

Set these up once, and your team stops debating format in every refinement session.

Step 2: Triage thin items with writing mode

Now you're ready for the actual pass. Open the first thin item — let's say it's titled "Fix login bug" with no description.

Click into Writing mode on the work item form. Pick the field you want to fill (Description), type a one-off prompt like "Expand this into a bug report based on the title. Note that we don't have repro steps yet — flag what we'd need to ask the reporter," and run it.

You get back a draft. It won't be right — the AI doesn't know your system. But it will give you a structured starting point that names the unknowns: "Reporter has not provided repro steps. To investigate, we need: browser/version, steps to reproduce, expected vs actual behavior, frequency."

Now your refinement work changes shape. Instead of writing a description from scratch, you're editing a draft and noting which questions to ask the reporter. That's a five-minute task instead of a twenty-minute one.

Writing mode is the right tool when the item is too unique for a template — bugs, weird one-offs, things where you need to think out loud with the AI rather than apply a standard format.

Step 3: Use Agentic AI Chat for the items you don't want to open

Some items in your backlog don't need rewriting — they need a decision. Is this a duplicate? Is it still relevant? Does it belong in this sprint or the next one?

Open the Agentic AI Chat in the Boards hub and ask in plain language: "Show me all bugs in the current iteration that haven't been updated in 30 days." Or: "Which items in the Sprint 24 backlog are missing acceptance criteria?" Or: "Find work items related to login that are still active."

The chat queries live work item data and answers. If you want it to make changes — close stale items, move them to a different iteration, add a tag — it proposes the changes and waits for you to review before anything is applied. You're not handing over the keys; you're getting a faster way to ask questions about your own backlog.

For a scrum master running a refinement session, this is the difference between "let me pull up that query" (cue twenty seconds of silence) and just asking a question while the conversation continues.

Step 4: Break down epics into child items in one pass

You've got an epic that needs to become five user stories. Or a story that needs to become eight tasks plus test cases. Normally this is the most tedious part of refinement: create child, type title, set type, link, repeat.

Open the work breakdown tab on the parent item. Pick the target type — Task, Test Case, User Story, whatever — and optionally add instructions like "Focus on the backend changes first, then the UI" or "Include test cases for the unhappy paths." The AI proposes a list of child items based on the parent's full context.

Then the part that matters: you review the list, toggle off the ones that don't fit, edit the titles inline, and create them all at once. You're not rubber-stamping — you're editing a proposal. But you skipped fifteen minutes of mechanical creation to get there.

This works especially well for generating test cases from a user story, because the parent already contains the acceptance criteria. The AI uses them as the basis for the test cases, and you adjust.

Step 5: Save the second pass for what only humans can do

Here's what I want to leave you with. The point of all this isn't that AI does refinement for you. The point is that the mechanical parts of refinement — formatting, structuring, decomposing, finding gaps — stop eating your Friday.

What's left is the part that matters: deciding what's actually important, talking to the people who reported the issues, making priority calls, figuring out which item is actually two items in disguise. That's the work product owners and scrum masters are good at. It's also the work that gets squeezed out when you spend the whole day writing.

Refinement done well isn't faster typing. It's more time for thinking.

Try it on your next refinement session

ClearSpecs AI installs on Azure DevOps directly from the Visual Studio Marketplace. The free tier gives you 10 prompts per month — enough to try the workflow above on a handful of work items and see whether it actually saves your team time.

If it does, the Starter plan is $10/month for 600 prompts, which covers most individual product owners. Teams typically land on the Business plan.

Install ClearSpecs AI on Azure DevOps

Or read the full documentation to see every feature in detail.