Write a Sprint Retrospective Prompt That Assigns Owned Actions

A sprint retrospective prompt has one failure mode that every template online walks right into: it produces themes nobody owns. You get a tidy "What went well / What didn't" summary, the meeting ends, and three weeks later the same problem shows up because no action item had a name attached.

Most retro content is a menu of formats. Mad-Sad-Glad, Start-Stop-Continue, sailboat. Those organize the conversation. They don't synthesize the output. The chained four-step frameworks floating around are mostly blog theory that nobody runs live during a 30-minute retro.

What a working team actually needs is one paste-ready prompt that takes the raw notes, ranks the themes, and assigns each action to a person with a due signal.

What you can do with a retro prompt

• Cluster a wall of sticky-note text into 4 to 6 ranked themes
• Separate signal from one-off venting so the team doesn't action a single grumpy comment
• Turn each top theme into an action item with an owner and a due signal
• Keep dissenting notes visible instead of averaging them into bland consensus
• Compare this sprint's themes against last sprint's to catch recurring problems
• Produce a short summary the team can paste straight into the next planning doc

Why "owned actions" is the whole job

Here's the take. A retro that ends with themes is half-finished. Themes are diagnosis. Action items with owners are treatment. Most AI retro prompts stop at diagnosis because it's the easy part, and the model is happy to write a thoughtful paragraph about "communication challenges" that obligates nobody.

The fix is a hard rule in the output contract: every action item must name an owner from {{team_roster}}, or the model has to flag that the theme has no clear owner. That second branch matters. An unowned action is a real finding, not a gap to paper over.

Anatomy of the prompt

Put the role and rules first, the raw notes in the middle, the contract last. A retro dumps a lot of unstructured text into {{retro_notes}}, and a contract placed before that wall of notes gets ignored. Models weight recent tokens, so the format spec belongs at the end.

Role: You are a retrospective facilitator.

Team roster: {{team_roster}}
Sprint: {{sprint_name}}
Prior themes (optional): {{prior_themes}}

Raw notes:
{{retro_notes}}

Output format:
1. 4-6 ranked themes, each with a one-line summary and a note count
2. Keep any dissenting note visible under its theme
3. Action items: Theme | Action | Owner (from roster) | Due signal
4. Mark any action with no clear owner as OWNER UNKNOWN
Return only the themes and the action table. No preamble.

How models cluster differently

Claude is steadier at honest clustering. It will keep a contradictory note ("standups are useless" next to "standups are the only sync we have") as genuine disagreement instead of collapsing both into "mixed feelings on standups." GPT-4o clusters fast but smooths conflict by default, so you have to say "preserve disagreement" explicitly. Gemini sometimes invents a sentiment label the notes don't support. The fix for all three is the same: tell the model to quote the note, not characterize it.

Step-by-step usage

1. Collect the notes

Export the retro board or the survey responses into one block. Don't clean them up. The mess is the input.

2. Set the roster

Fill {{team_roster}} with the names that can own an action. The model can only assign owners it's been given.

3. Run the prompt

Paste it. Read the ranked themes first, then check the action table for OWNER UNKNOWN rows.

4. Resolve the unowned actions

Every OWNER UNKNOWN row is a live decision for the team. Assign it in the room, then update the prompt input and rerun if you want a clean record.

5. Carry themes forward

Save this sprint's themes into {{prior_themes}} for next time. That's how you catch the problem that keeps coming back.

Prompt-craft patterns

Pattern one: rank by note count, not by drama. Tell the model to weight a theme by how many notes mention it. One loud complaint shouldn't outrank a quiet pattern five people noticed.

Rank themes by the number of distinct notes that mention them,
not by how strongly any single note is phrased.

Pattern two: quote, don't summarize, the dissent. A summarized disagreement reads as consensus. A quoted one stays sharp. This is the single most useful guardrail in the whole prompt.

Pattern three: a due signal, not a date. The model doesn't know your calendar, so don't ask for a date it'll hallucinate. Ask for "this sprint / next sprint / backlog" and let the owner set the real deadline.

Variables you'll set

Getting started

1. Export the retro board into {{retro_notes}} as-is.
2. List the team in {{team_roster}}.
3. Run the prompt and scan the ranked themes.
4. Find every OWNER UNKNOWN row and assign it live.
5. Confirm no theme invented a sentiment the notes don't back.
6. Paste the action table into your tracker.
7. Stash the themes in {{prior_themes}} for next sprint.

The Sprint Retrospective Agent Pack ships this with the dissent-preserving rule and the owner-assignment contract already wired in, so you're not rebuilding the prompt every two weeks.

When you run retros every sprint

The retro feeds the next planning cycle, so it pairs naturally with the product roadmap prompt that turns those learnings into a ranked plan. And if your board is cluttered enough that the retro keeps surfacing "too much WIP," the Jira backlog cleanup prompt clears it before the next sprint, and the Backlog Cleanup Agent Pack ships the verdict contract that does it. For turning ranked retro themes into a committed plan, the Connected Roadmap Alignment Agent Pack picks up where the retro leaves off.