Analyze Sprint Meeting Minutes with Claude: Copy-Paste Prompts for Planning Poker Data

To analyze sprint meeting minutes with AI, export your planning poker session as a text or Markdown file, add a short list of which round was which story, and paste it into Claude (or another assistant your organization has approved) with a specific question, such as "which stories had the widest vote spread and why might that be?" Specific prompts get useful answers. "Analyze this" gets a generic summary.

This guide gives you eight prompts you can copy, a sample anonymized minutes file so you can see what the AI is working with, the shape of a good answer, and the privacy checks to run before you paste anything.

What the AI can and can't see

Be clear about the workflow, because it's easy to overstate. Alignlee exports meeting minutes as a Markdown (.md) file. Alignlee doesn't call any AI service. You open the file, check it, and paste it into an assistant yourself. The same approach works with notes from any tool, as long as they're text.

An Alignlee planning poker export contains:

• Session name, date, participant names, and voter count
• For each round: the time, each person's vote and confidence percentage, the confidence-weighted result, the most common vote, and average confidence
• Any Confidence Pulse rounds (team sprint confidence on a 1 to 5 scale) with the spread
• A session summary table

It does not contain story titles, acceptance criteria, or what people said. Rounds are labeled "Round 1", "Round 2", and so on. That means the AI can tell you Round 4 had votes from 3 to 13 with low confidence, but it can't tell you why unless you add context. Two minutes of annotation makes the analysis much more useful.

Add a context block before you paste

Put this at the top of the file (or the start of your message):

CONTEXT
Team: 6 developers, 2-week sprints, Fibonacci scale, velocity last 3 sprints: 31, 27, 34
Round 1 = Story A: Export invoices as PDF
Round 2 = Story B: Retry failed webhook deliveries
Round 3 = Story C: Audit log for role changes
Round 4 = Story D: Migrate reports to new data warehouse
Round 5 = Story E: Rate-limit public search API
Notes: Round 4 discussion was about unknown schema differences. Priya out next sprint.

Notice the generic labels ("Story A") and no customer names. More on that in the privacy section below.

A sample minutes snippet (anonymized)

Here is a trimmed example of what an exported session looks like, with names replaced by roles:

# Planning Poker Session - Meeting Minutes

**Session:** Sprint 24 Refinement
**Participants:** Dev1, Dev2, Dev3, Dev4, QA1 (5 voters)

## Planning Poker Decisions

### Round 3 - Story Estimate
**Final Estimate:** 5.2

| Participant | Vote | Confidence |
|-------------|------|------------|
| Dev1 | **5** | 80% |
| Dev2 | **5** | 75% |
| Dev3 | **8** | 60% |
| Dev4 | **5** | 85% |
| QA1  | **3** | 50% |

**Statistics:**
- Most common vote: **5** (3 voters)
- Average confidence: **70%**

### Round 4 - Story Estimate
**Final Estimate:** 7.9

| Participant | Vote | Confidence |
|-------------|------|------------|
| Dev1 | **3** | 30% |
| Dev2 | **13** | 40% |
| Dev3 | **8** | 35% |
| Dev4 | **8** | 45% |
| QA1  | **5** | 25% |

**Statistics:**
- Most common vote: **8** (2 voters)
- Average confidence: **35%**

## Confidence Pulse Decisions
### Pulse 1 - Sprint Confidence Check
**Team Confidence Score:** 3.2/5.0
**Vote Spread:** 3 points (2-5)

The "Final Estimate" is confidence-weighted: low-confidence votes count for less. Round 4 tells you a lot even without notes. A 3 to 13 spread with average confidence of 35% is a story the team doesn't understand yet, regardless of what number they agreed on.

Eight copy-paste prompts

Each prompt assumes you've pasted the context block plus one or more exported sessions. Replace the bracketed parts.

1. Find estimation outliers

Below are planning poker minutes with a context block mapping rounds to stories.
For each round, calculate the spread (highest minus lowest numeric vote, ignoring "?")
and the average confidence. List the rounds where the spread covers 3 or more
cards on the Fibonacci scale OR average confidence is below 50%. For each, say
which participants were at the extremes. Don't speculate about reasons I haven't
given you; flag where you'd need more information.

2. Spot stories that need splitting

Using these minutes, identify stories that look too big or too uncertain to bring
into the sprint as-is: any story with a result above 8, or a wide spread combined
with low confidence. For each one, suggest 2-3 ways it could be split into
smaller vertical slices, using only the story description in my context block.
Mark each suggestion as "needs team validation".

3. Summarize risks for sprint planning

Read the minutes and the notes in the context block. Produce a risk list for
sprint planning with columns: Story | Signal in the data | Likely risk | Question
to ask the team. Base "signal in the data" only on votes, confidence, and my notes.
Keep it to the top 5 risks.

4. Draft a sprint goal

Here are the stories we estimated and expect to pull into the next sprint, with
their agreed points. Our recent velocity is [31, 27, 34]. Draft three alternative
sprint goals, each one sentence, that describe the outcome for users rather than
listing stories. For each, list which stories support it and which don't fit.

5. Check the commitment against capacity

Our planned sprint total is [36] points. Velocity over the last three sprints was
[31, 27, 34]. One developer is out for [3 of 10] days. Using the minutes, tell me
which stories carry the most estimation risk (low confidence or wide spread) and
which 1-2 stories would be the most sensible to drop if we want a realistic plan.
Show your arithmetic.

6. Prepare questions for the Product Owner

For every round where average confidence was below 50%, write 2-3 specific
clarifying questions the Product Owner could answer before our next session.
Base them on the story description in my context block. Avoid generic questions
like "what are the requirements?"

7. Compare sessions over time

I'm pasting minutes from [4] refinement sessions over the last [2] months, each with
a context block. Compare them: Is average confidence trending up or down? Is the
spread of votes narrowing? Are the same kinds of story (by my descriptions)
repeatedly producing wide spreads? Present a small table per session, then
3 observations. State clearly where the sample is too small to conclude anything.

8. Build a reference-story list for calibration

From these sessions, pick one story the team agreed on with high confidence at each
of the values 1, 2, 3, 5, and 8. For each, give the story description and one line
on why it's a good reference. If there's no strong example at a value, say so
rather than picking a weak one.

Prompt 8 pairs well with the process in calibrating story points, and prompt 2 with estimating large epics.

What a good answer looks like

Ask for structure and you'll get something you can scan in a minute. For the sample above, a useful response to prompt 1 would look roughly like this:

Outliers

| Round | Story | Spread | Avg confidence | Extremes |
|-------|-------|--------|----------------|----------|
| 4 | D: Migrate reports | 3 to 13 (4 cards) | 35% | Dev1 (3), Dev2 (13) |
| 3 | C: Audit log | 3 to 8 (3 cards) | 70% | QA1 (3), Dev3 (8) |

Notes
- Round 4: spread and confidence both suggest the team lacks shared
  understanding. Your notes mention unknown schema differences; a spike
  may be worth more than a re-vote.
- Round 3: confidence is fine; the disagreement may be about test scope
  (QA voted lowest). Worth a 2-minute check.
- Need more info: I can't tell whether Round 4's agreed 8 was re-voted.

Warning signs of a bad answer: invented reasons for disagreements you didn't describe, precise-sounding percentages from a single session, or confident claims about "the team's accuracy" when the file contains no actual delivery data. If you see these, reply "only use what's in the minutes and my notes" and ask again.

A simple routine that fits into a sprint

Save the files somewhere consistent, for example team-notes/2026-10-07-refinement.md, so you can paste several at once later. The AI output is a starting point for a team conversation. Share it with the team, and let the team decide what to act on. Planning poker is useful because people talk through their assumptions; the AI is only a quicker way to notice which conversations you still need to have.

Using AI during the session (carefully)

Some facilitators keep an assistant open during refinement. When votes come back 2 to 13, they paste the story text and ask for clarifying questions about where people might be interpreting the scope differently. That can help a stuck discussion, but keep two rules:

1. Never ask the AI for the estimate. It doesn't know your codebase, your test setup, or your team. If you start anchoring on its number, you've lost the reason for voting independently (see stopping anchor bias in planning poker).
2. Keep it under two minutes. If the team needs more than that, the story needs refinement, not a better prompt.

Privacy cautions before you paste

Meeting minutes look harmless, but they often contain things your organization would not want sent to an outside service.

• Check your policy first. Only use AI tools your organization has approved for work data. Many companies have an approved enterprise assistant and a ban on consumer accounts for work content. If you work on government or regulated projects, assume external tools are off-limits until someone tells you otherwise.
• Replace people's names. The export includes participant names next to votes. Swap them for roles (Dev1, QA1). Analyzing who votes high or low by name can feel like performance monitoring to the team, even when it isn't meant that way.
• Use generic story labels. "Story D: migrate reports" rather than the customer, product codename, or ticket summary that reveals a client.
• Strip anything secret. No credentials, internal URLs, security vulnerabilities, revenue figures, or unreleased product plans.
• Know the retention rules. Understand whether your assistant keeps conversations and whether they can be used for training under your plan, and delete conversations you don't need.
• Tell the team. If you're pasting their votes into an AI tool, say so and explain why. It's their data too.

FAQ

Can Claude analyze planning poker results?

Yes, if you give it the results as text. Export the session, add a short note mapping each round to a story, and ask a specific question such as which rounds had the widest spread or lowest confidence. Claude works from what you paste; it can't see your tracker or your codebase.

Does Alignlee have built-in AI analysis?

No. Alignlee exports meeting minutes as a Markdown file and doesn't send anything to an AI service. If you want AI analysis, you paste the file into an assistant yourself, which also means you decide what gets shared.

What should I remove before pasting meeting minutes into an AI tool?

Replace participant names with roles, swap customer or product names for generic labels, and remove anything confidential such as credentials, security issues, or financial figures. Above all, confirm the tool is approved by your organization for work data.

How many sessions should I analyze at once?

One session is enough to find outliers and draft questions for the Product Owner. For trends, use at least three or four sessions, and treat any conclusions as hypotheses for the team to check, since a handful of sessions is a small sample.

Can AI estimate my stories for me?

It can produce a number, but you shouldn't use it. Estimates are valuable because the people doing the work surface different assumptions. Use AI to find where those assumptions diverge, then let the team vote.