Run a team retrospective

How SquadBear's retro board moves through its stages, how blind brainstorm and voting keep it honest, and how the check-in scorecard closes the loop on last time's actions

A retrospective is a structured, live conversation about how the last stretch of work actually went. SquadBear runs it as a shared board every participant opens at once, on their own device, moving through the same stages together. Like a health check, it is private by design and ends in a small number of owned actions — not a wall of sticky notes nobody revisits.

Where it lives

Open Team Health → Retrospectives in the sidebar. The list groups sessions into Open (a live board with its stage and participation), Upcoming drafts, and Closed (with a participation count and an AI summary). New retro opens the wizard: pick a Team, a Format, an optional Title, an Attribution mode, Votes per person, an optional icebreaker question, and the blind toggles. Only a team's lead, its manager, or an admin can start one; everyone invited joins the same board once it opens.

Formats are the classic column layouts — Start / Stop / Continue, Mad / Sad / Glad, 4Ls, Sailboat, Starfish, and a lightweight Team Health Check. The columns you pick decide the prompts people write against; the mechanics below are the same for all of them.

The stage flow

A retro walks one hop at a time through a fixed sequence. The facilitator moves everyone forward with Next, and can step back a single stage if the room needs it. Closing is a separate, deliberate step at the end — never a stage you "advance" into by accident.

Check-in — the team looks back before it looks forward (see the scorecard below) and everyone answers a one-line icebreaker.
Brainstorm — people add cards to the columns. This is where blind brainstorm matters.
Group — cards are dragged together into themes; the facilitator can rename or delete groups and ask the AI for merge suggestions.
Vote — each person spends a fixed budget of votes on the cards and groups that matter most.
Discuss — tallies are revealed, the top topics rise to the top, and the team talks through them, leaving comments and drafting actions.
Review — a last look at the actions (aim for one or two, each owned and dated) before closing.

While brainstorming, grouping or voting, anyone can flag themselves Ready — a lightweight, private signal that only tells the facilitator who's still working, with no effect on the board itself. It clears automatically for everyone the moment the stage moves on.

Stepping back one stage is safe and keeps everything people wrote. Only Close is irreversible.

Two toggles keep the retro honest, and both are on by default:

Blind brainstorm hides everyone else's cards while people are still writing, so the loudest voice or the most senior person can't anchor the room. You always see your own cards; for everyone else's you see only a per-column count ("+3 hidden"). The moment the team reaches the group stage, every card is revealed at once.
Blind voting hides the running tallies while voting is open, so nobody piles onto an early leader. You always see your own dots and your remaining budget; the aggregated tallies appear for everyone the instant the team reaches discuss.

The blindness is enforced on the server, in a single projection every surface reads through — the board, the REST API, the MCP tools and the live socket all get the same hidden frame. A card's text, its id, and its author simply are not sent to other participants until the stage that reveals them. Turn either toggle off at creation time if you'd rather run the retro fully open.

Attribution modes

Cards carry attribution according to the Attribution mode chosen at creation:

Named shows the author's real name.
Alias gives each person a stable, playful pseudonym for the session ("Amber Otter"), so you can tell two cards came from the same person without knowing who.
Anonymous attaches no author at all.

Votes are always anonymous, in every mode — the server only ever stores an aggregate tally, never who voted for what.

The check-in scorecard — "you said, we did"

Check-in is where a retro earns its keep. Instead of starting from a blank page, the board opens with the team's open improvement actions — the ones created in earlier retros or health checks. Walk the list, and mark anything that's genuinely done with Mark done right there on the board (it completes the same action that shows up on the workspace-wide Actions list). It's the loop that makes the next retro worth running: last time's promises are the first thing everyone sees.

The icebreaker question itself isn't fixed either — the facilitator can pick a different one, edit it, or turn it off entirely, right up until check-in ends. Once brainstorming starts, it's locked for the rest of the retro.

Vote budgets and stacked votes

Each participant gets a fixed number of votes — Votes per person. Unless the facilitator pins that number when creating the retro, it's derived automatically the moment voting opens: roughly one vote per three cards and groups on the board, clamped between 3 and 10, so a sprawling board earns more votes than a quiet one. The facilitator can still nudge the number up or down while voting is open, from 1 to 20, if the room needs it. By default a person can back a given card or group only once, which spreads attention across topics. Turn on Stacked votes to let people pile several of their votes onto a single item, so intensity shows through, not just breadth. Either way, the budget is the cap: your remaining votes tick down as you spend them, and retracting a vote hands it straight back.

Grouping and AI merge suggestions

In the group stage, drag one card onto another to form a theme, drop a card on a group's header to add it, or drop it on the column background to pull it back out. There's a ⋯ menu on every card for the same moves by keyboard or on touch. If the workspace has AI enabled, the facilitator can ask for merge suggestions: the model reads the card texts, proposes pairs that say the same thing, and each suggestion applies in one click. Nothing is grouped without a human accepting it.

Discuss: the shared spotlight

The moment the retro reaches discuss, tallies are revealed and the facilitator drives a shared spotlight: whatever card or group they focus lights up on every open board at once, so the room looks at the same topic together instead of scattering. Working top-voted first, the facilitator marks each topic discussed as the conversation moves on, checking it off a shared walk-through list. The spotlight itself clears the moment the retro leaves discuss — the checked-off list stays.

Closing a retro

Closing freezes the retro for good, in one deliberate, confirmed step:

The results are snapshotted — the tallies, the participation count, the column counts and the ranked ordering are frozen into the session so post-close reads never depend on the live data again.
Cards are irreversibly anonymized: every author hash is rotated to a fresh random value, so a contribution can never be traced back to a person, even by someone with raw database access.
The votes are deleted — the frozen tallies are the record from then on.

There is no undo. A retro left running for more than a week auto-closes the same way, so an abandoned board never keeps its anonymity key alive indefinitely. After close, the board becomes a read-only record with an AI-generated summary of the discussion.

Worked example

Mila runs a Start / Stop / Continue retro for the Platform team, alias mode, both blind toggles on, five votes each. Check-in opens with two actions from last sprint — the team marks the on-call rebalancing one done. In brainstorm, everyone drops cards blind, seeing only counts climb. At group, the cards reveal and three "too many meetings" cards merge into one theme (the AI flagged the pair; Mila accepted). Voting is blind; when the team hits discuss, "Meeting overload" leads with nine votes. They draft one action — "Cut the Weds sync to 30 min," owner Mila, due in a week — review it, and close. The votes vanish, the aliases dissolve, and the summary lands on the closed board.

An agent facilitates a retro through the same MCP tools the board uses: create_retro_session and set_retro_stage to run the flow, add_retro_cards to bring topics in (for example from a tracker), group_retro_cards and cast_retro_vote to organize and prioritize, update_retro_session to override the auto-derived vote budget, change the icebreaker or run the shared countdown mid-retro, set_retro_discuss_focus to drive the shared spotlight through discuss (top-voted topics first, marking each discussed as it wraps up), suggest_retro_actions to propose next steps (read-only — recording still goes through create_action after you confirm), close_retro_session to freeze it, and summarize_retro_session for the closed write-up. get_retro_session polls the live state, and always respects the blind rules — an agent never sees hidden card text or tallies before the stage that reveals them.