AI theme clustering
Theme clustering is the agent reading the whole board and grouping cards that say the same thing under a short shared label. It replaces the manual grouping step that used to sit between writing and voting.
There is no dragging cards onto each other to group them. Grouping is something you ask for and get back, not something you do by hand.
What it does
Section titled “What it does”The facilitator selects Cluster in the dock. The board’s cards go to the model, which returns a set of themes — each a two-to-four-word label and the cards it covers. Every clustered card picks up a sparkle chip carrying its theme name.
Three rules shape the result:
- A theme needs at least two cards. A single card is not a theme.
- Cards that share no topic are left alone. No chip, no forced group.
- It groups, it doesn’t rewrite. Cards keep their own text, author, column, position, and votes. The theme is a label laid over the board, not a merge that destroys anything.
The run happens off the request path, so the board stays usable while it works. The Cluster button spins, and the themes appear on every open client at once when it lands — nobody needs to refresh.
Why before the vote
Section titled “Why before the vote”Because duplicates split votes.
Five cards saying “deploys are painful” in five different ways compete with each other. On a five-vote budget, the room’s clearest shared frustration can finish behind a snappier one-off, and the retro spends its discussion time on the better-phrased issue rather than the more important one.
Clustering makes that visible before anyone spends a vote. It’s why the board offers it proactively: when a guided board enters Vote with six or more un-clustered cards, the facilitator gets a strip suggesting it.
What the model can see
Section titled “What the model can see”Card text and column title. That’s the whole input.
No author, no participant identity, no display name, no vote count — the same guarantee the board’s own wire enforces for hidden and anonymous feedback. A card written on an anonymous board is exactly as anonymous to the model as it is to the room.
Re-running replaces what you have
Section titled “Re-running replaces what you have”This is the part worth knowing before you press the button a second time.
Clustering is a whole-board operation: it deletes the board’s existing themes and inserts the new set. It does not add to what’s there, and it does not preserve a theme just because it looks the same.
Because of that, a second run on a board that already has themes opens a confirmation naming how many will be replaced, and says plainly that it can’t be undone.
When it’s offered
Section titled “When it’s offered”The Cluster control appears only for the facilitator of a guided board — the phase structure is what makes “before the vote” a real moment.
It also depends on the board’s own Theme clustering toggle, set when the board was created alongside the other two agent features (trends and probe suggestions). All three default to on.
What it costs
Section titled “What it costs”Retro boards run free on every plan. The agent that reads them does not: theme clustering, cross-retro trends, and probe suggestions all sit behind the retro AI entitlement, which is unavailable on Free and unlimited on Pro, Business, and Enterprise.
On a Free workspace the Cluster button still appears, marked with a small lock. Selecting it opens a panel explaining the three agent features rather than running anything. Nothing else about the board changes — the phases, the votes, the spotlight, and the summary all work as normal. See Plans and tiers.
When it fails
Section titled “When it fails”If the model can’t produce a usable grouping, the run stops and tells you so rather than retrying quietly. The board is untouched — press Cluster again if you want another attempt.
A board where genuinely no two cards share a topic returns nothing, and says so. That’s a real answer about your board, not a failure.
What themes feed afterwards
Section titled “What themes feed afterwards”Themes are the unit the rest of the retro AI works in:
- Trends across retros on the closed retro’s summary track which themes repeat and grow, run over run.
- Probe deeper offers to draft an Agent Survey that interviews the team about the top theme, async.
- Ask can be pointed at a closed retro to query its themes in context — see Ask overview.
A board that was never clustered simply has no themes for any of that to work from.