Retrospective best practices

The mechanics of running a retro are in Running a retro. This is about what makes the hour worth spending.

Run a continuous board

The single highest-leverage habit: create the retro at the start of the sprint, not five minutes before the meeting, and let people add cards as things happen.

The alternative is asking a room to remember two weeks on demand. What you get back is the last three days, plus whatever was loud. Everything from the first week — including the thing that actually caused the mess — is gone.

A continuous board also changes the tone. A card written the afternoon someone lost two hours to a flaky test is specific and factual. The same card reconstructed a fortnight later is a generalisation about testing.

Set the board to always-visible for this — turn Private until reveal off in Settings. Privacy is there to stop anchoring in a live writing sprint; on a board people top up over two weeks it just means nobody can see what’s already been said, and the same card gets written four times.

Decide the privacy posture deliberately

Private until reveal is on by default, and for a live session it earns its place: cards that appear as they’re typed converge, and the first theme of the first ninety seconds becomes the theme of the retro.

Anonymous is different, and worth being honest about. It’s the right call when attribution would change what gets written — a post-incident review, a room where someone has been burned before, a retro with a manager in it. It’s the wrong call as a standing default: it costs you every follow-up question, and a team that can only be honest anonymously has a problem the retro format won’t fix.

Both settings are covered in Private feedback and anonymity.

Timebox the writing, not the discussion

Put a 5-minute timer on the Feedback phase. Everyone sees the same countdown, and it does the “right, let’s move on” for you.

Don’t timer the Review. A conversation with a visible clock on it becomes a conversation about the clock. If Review is running long, the cause is upstream — too many cards, or a vote that didn’t discriminate — and the fix is next time’s board, not this one’s stopwatch.

Cluster before you vote, not after

Duplicates split votes. Five cards saying “deploys are painful” in five different ways compete with each other, and on a five-vote budget the team’s clearest shared frustration can finish behind a snappier one-off.

Merging them first is the difference between voting on issues and voting on phrasing. On a paid plan the board offers this proactively when you enter Vote with a full board — see AI theme clustering. Without it, read the duplicates out loud before voting and let the room consolidate.

Discuss three things, not thirteen

Stop when the vote counts flatten. A run of cards on one vote each is the board telling you that the room has finished prioritising and started listing.

Three well-discussed cards that produce two changes beat thirteen cards that produce a list nobody looks at. The discussion list dims what you’ve done and shows how far down you are — use it to decide when to stop, not to feel behind.

Write the outcome down in the retro

Whatever the team decides has to go in the comment thread on the card it came from. Nothing else in the session persists — the discussion isn’t recorded, and the summary carries cards, votes, and comment counts, not what was said out loud.

A decision that only exists in the room did not happen. This is the most common way a good retro produces nothing.

Start the next retro by reading the last one

Open the previous retro’s summary before the new board goes live and spend two minutes on the top-voted cards. Did anything change?

This is the whole feedback loop. A team that never revisits its last retro is running a well-facilitated venting session, and people work that out quickly — you’ll see it in the card counts before anyone says it.

On a paid plan the summary’s Across retros panel does part of this for you, showing which themes keep coming back and growing.

Watch the temperature, gently

The temperature check is worth reading as a series, not a number. Your team’s baseline says more about its calibration than its health; the movement across three or four retros is the signal.

Watch the answer count too. Nine out of ten is a reading. Three out of ten is a different finding, and usually a more urgent one.

Don’t turn it into a target. A temperature that becomes a metric to improve becomes a metric people manage, and then it stops telling you anything.

Change the format when the retro goes stale

Running the same three columns every sprint for a year produces the same three columns’ worth of thinking. The format is a set of prompts, and prompts wear out.

When the board starts feeling rote, switch: Sailboat pulls the conversation forward into risks, 4 Ls goes deeper for a quarter or project end, Incident post-mortem is timeline-first and blameless. Twenty formats ship — see Retro formats.

Keep the room small enough to talk

A retro is a conversation. Past roughly ten people it becomes a broadcast with a queue, the quiet people stay quiet, and the vote spreads too thin to discriminate.

If the group is genuinely larger, run separate retros per team and use an Agent Survey to reach across them — that’s the tool built for asking many people the same thing.