What is backlog refinement?
Backlog refinement is how a team turns rough backlog items into work it can start. The 2020 Scrum Guide defines it this way: "Product Backlog refinement is the act of breaking down and further defining Product Backlog items into smaller more precise items. This is an ongoing activity to add details, such as a description, order, and size."
The Agile Alliance glossary describes the same thing from the team's side: the product owner and some, or all, of the rest of the team "review items on the backlog to ensure the backlog contains the appropriate items, that they are prioritized, and that the items at the top of the backlog are ready for delivery." It lists the activities that make up a session:
- removing user stories that no longer appear relevant;
- creating new user stories in response to newly discovered needs;
- re-assessing the relative priority of stories;
- assigning estimates to stories that have yet to receive one;
- correcting estimates in light of newly discovered information;
- splitting user stories that are high priority but too coarse-grained to fit in an upcoming iteration.
The output is a backlog whose top items are clear, small and in order. Lower items can stay rough until they move up.
Backlog refinement vs backlog grooming: is there a difference?
No. Backlog grooming is the older name for the same practice. The Agile Alliance entry says that "due to the increasingly negative connotation of the word grooming", the activity is increasingly known as backlog refinement, and that grooming was meant to suggest "trimming, pruning, and cleaning, as with a plant." The name changed in the Scrum Guide itself, as its revision history shows:
| Year | What happened | Source |
|---|---|---|
| 2005 | Earliest recorded use of "backlog grooming", by Mike Cohn on the Scrum development mailing list | Agile Alliance glossary |
| 2008 | Kane Mar gives one of the first formal descriptions, as a regular meeting called "Story Time" | Agile Alliance glossary |
| 2011 | The Scrum Guide adds "the practice of Product Backlog Grooming" | Scrum Guide revisions |
| 2013 | "The Product Backlog is refined rather than groomed" | Scrum Guide revisions |
| 2020 | Current guide: refinement is "an ongoing activity"; the 10 percent capacity guideline from 2017 is gone | Scrum Guide 2020 and 2017 |
You will still meet "grooming" in older tools and articles. Treat the two words as the same thing.
Is backlog refinement a Scrum event?
No. Refinement is not one of the Scrum events. The 2020 guide has the Sprint, which it calls "a container for all other events", and four events inside it: Sprint Planning, Daily Scrum, Sprint Review and Sprint Retrospective. Refinement is an activity that happens when it is needed. During the Sprint, the guide says, "The Product Backlog is refined as needed", and in Sprint Planning "The Scrum Team may refine these items during this process."
The guide no longer says how much time to spend. The 2017 edition said refinement "usually consumes no more than 10% of the capacity of the Development Team"; the 2020 edition dropped that line. The Agile Alliance glossary says refinement "may be an officially scheduled meeting or an ongoing activity", so the format and the timebox are the team's call.
Two roles matter. "The Developers who will be doing the work are responsible for the sizing." And the Product Owner is accountable for the backlog, including "Ordering Product Backlog items". The Product Owner "may do the above work or may delegate the responsibility to others. Regardless, the Product Owner remains accountable." That last sentence is the rule to keep in mind when an AI agent joins in.
How do you run a backlog refinement session?
Work from the top of the backlog down, and stop when the next few items are ready. These six steps cover the Agile Alliance activities in a sensible order:
- Prune. Remove items that no longer matter and merge duplicates. A shorter list is easier to order.
- Order. The Product Owner puts the remaining items in order of value, so the team refines what comes next first.
- Clarify the top items. For each one, write the goal and the acceptance criteria, and note dependencies and open questions.
- Split what is too big. Break any item the team can't finish in one Sprint, or review in one sitting, into smaller items that each deliver something. Our post on breaking work down for AI agents covers how far to go.
- Size. The people who will do the work estimate it, in whatever unit the team uses.
- Check against your definition of ready. Items that pass are ready for selection. If your team keeps a definition of ready, this is where you apply it.
As a common practice, new items that come up along the way go to the bottom of the backlog unless the Product Owner orders them higher.
What changes when AI agents work the cards?
The refined card becomes the agent's whole brief, so refinement has to be more complete. A teammate who picks up a vague item can ask a question in stand-up. An agent working alone reads the card and starts. Anything the session left out, the agent guesses.
That changes what "clear enough" means. For a card an agent will work, the description needs the goal, a check the agent can run for each criterion, where to look and what to leave alone, and who reviews the result. A person can carry some of that in their head. An agent can't.
Agents can also help with the refinement itself. Anthropic's Claude Code best practices suggest that when you don't yet know exactly what you want, you ask Claude to interview you, for example: "Interview me until you could write the spec, then write it to SPEC.md." Per the same page, Claude replies with multiple-choice questions about the parts you haven't pinned down. That is refinement of a single item, done as a conversation. Use it on the vaguest card in the backlog.
How should AI agents and people share refinement?
Agents prepare, people decide. An agent can read every card and check each one for missing acceptance criteria. But ordering and accepting work are decisions the Product Owner stays accountable for, and sizing belongs to whoever answers for the result. Mapped onto the Agile Alliance list:
| Refinement activity | An agent can prepare | A person decides |
|---|---|---|
| Remove items that no longer matter | Flag cards that look stale or duplicated | Archive or delete them |
| Create items for new needs | Draft the card text in a comment | Whether it belongs in the backlog |
| Re-assess priority | Point out dependencies that affect the order | The order (the Product Owner) |
| Estimate and correct estimates | Note what makes a card big: unknowns, many files, other teams | The size |
| Split coarse items | Suggest a split into smaller cards | Which pieces to create |
| Get top items ready | Check each description against your definition of ready | Moving the card to Ready |
How do you run an agent refinement pass on a Flux board?
Have the agent read the Backlog column, comment what each card is missing, and leave every change to a person. Use a board with Backlog, Ready, In Progress, Needs review and Done. Connect the agent through the Flux MCP server, then give it this pass in CLAUDE.md, AGENTS.md or a skill:
## Backlog refinement pass (prepare only, a person decides) 1. Find the Backlog column ID with get_board, then read the column with list_cards. Work from the top down. 2. For each card, read it with get_card. Skip it if your last refinement note has no reply yet. 3. Check the description for: - Goal: the outcome, in one or two sentences. - Acceptance criteria, each with a check someone can run. - Scope: where to look and what to leave alone. - Depends on: card IDs, or "none". 4. Look for likely duplicates with search_cards (it matches titles only). 5. If anything is missing, or the card looks too big for one review, post ONE comment (add_comment takes the card's UUID, not the SHOP-57 style ID) that starts "[refine-bot] Refinement notes:" and lists: - each gap, with suggested wording; - a suggested split into smaller cards, if it is too big; - possible duplicates, by card ID. 6. Do not edit titles or descriptions, change priority, reorder or move cards. A person applies the changes and moves cards to Ready. 7. Finish with a list of the cards you commented on and the cards that already look ready.
Here is the kind of comment it leaves on a vague card:
[refine-bot] Refinement notes: - Goal is unclear: "export" could mean a CSV download or a full account data request. Which one? - No acceptance criteria. Suggest: "Clicking Export on the Orders page downloads a CSV with one row per order; tests/export.spec.ts passes." - Too big for one review. Suggested split: 1. CSV export endpoint and its test 2. Export button on the Orders page - Possible duplicate: SHOP-57 "CSV download for orders". Nothing on this card was changed.
The person who owns the backlog reads the notes, edits the card, answers the questions in a reply and moves the card to Ready when it passes. The pass is written this way for these reasons, checked against the Flux server code in October 2026:
- One call reads the whole column. The MCP tool
list_cardsreturns every card in a column, in board order, with its title, description, priority and due date. Comments come only fromget_card, card by card. - Keep criteria in the description.
get_cardreturns the card's own fields and its comments, but not its checklist items or assignees, so anything kept there is invisible to an MCP agent. - Duplicate search is title-only.
search_cardsmatches card titles only and returns at most 20 results, from every board in your workspaces, archived cards included, so it finds "CSV download for orders" but not a card that only mentions CSV in its description. - Comment, don't rewrite. Changes made over MCP are not written to the Flux activity log, so if the agent rewrote a description with
update_card, there would be no before-and-after record to check or undo. A comment leaves the original text as it was. - "Prepare only" is an instruction, not a permission.
add_commentneeds a Write-scope key whose owner is at least a workspace member, and that same key can also edit, move and reorder cards. Nothing in Flux limits a Write key to comments, so the limit lives in the instructions. - Pruning stays with people. No Flux MCP tool archives or deletes a card, so stale cards are removed in the app.
- Sign the notes. Comments made with an API key appear under the name of the person who created the key, and
get_cardreturns comments without author names. The "[refine-bot]" prefix tells people and later runs who wrote it.
For splitting, people in the Flux app can also use the Split ticket button on a card. It suggests 2 to 5 smaller cards from the card's title, description and checklist. You edit the list before anything is created, and Flux then adds the chosen cards to the same column, linked to the original. It works in a signed-in browser session only, not with API keys. The work breakdown post covers it in more detail.