AI for Product Managers: Turn Scattered Feedback Into a Weekly Synthesis

Feedback arrives everywhere: a #customer-feedback channel, support threads, sales call recordings, the occasional email forwarded with "thought you'd want to see this". Product managers spend hours a week re-reading it to find the themes. This recipe gives that job to an agent that reads all of those sources on a schedule and delivers a synthesis with evidence, so your time goes to deciding, not collecting.

What you get

Every Monday morning, a thread from @product with:

1. The week's feedback grouped into themes, each with a count, three representative quotes, and links back to the source messages, threads or transcripts.
2. Which themes are new versus recurring, compared with last week's synthesis saved in your drive.
3. An offer to draft a one-page spec for the top theme.

Step 1: Connect the sources

In Integrations, connect the places feedback lives. This recipe uses:

• Slack for the feedback channel and anything customers say in shared channels.
• Gmail for support or founder inboxes where feedback lands.
• Fireflies for call transcripts (sales and customer calls).

All three are read sources here. Damon classifies every discovered action as a read or a write, and reads are auto-approved by default, so the agent can gather freely while anything that changes state still asks.

Step 2: Create the product agent

• Name: Product, @product
• Personality: "Evidence first. Quote customers verbatim. Distinguish what customers say from what they might mean, and label the second as inference."
• Goals: "Make sure no recurring customer problem goes unnoticed for more than a week. Keep the team looking at the same evidence."
• Integrations: Slack, Gmail, Fireflies.

Step 3: Run the synthesis once, by hand

In chat with @product:

Gather customer feedback from the last 7 days: everything in #customer-feedback, support threads in Gmail labeled "Feedback", and Fireflies transcripts of calls tagged customer. Group into themes. For each theme: a name, how many distinct customers raised it, three verbatim quotes with links, and whether it's about a bug, a missing capability, or confusion. Save the result to Drive as "Feedback/2026-W39.md".

Check that the grouping matches your judgment. If it's splitting one theme into three, or merging two, say so; then have it save your grouping preferences and the file naming as a skill so the next run follows them.

Saving to the drive is a file.write action; approve it the first time, then auto-approve it for this workflow. The saved files are what make week-over-week comparison possible.

Step 4: Schedule it and add the spec draft

Ask: "Run this every Monday at 7am, compare with last week's file, message me the synthesis, and offer to draft a spec for the top theme." The workflow:

1. Trigger — every week, Monday, 07:00.
2. Agent step — gathers, groups, compares with last week's file in Drive, writes this week's file.
3. Message step — delivers the synthesis and the offer.

When you say yes to the spec, the agent drafts it in Drive from the evidence it already has: problem statement, who's affected (with the quotes), what customers are doing today instead, and open questions. Editing a draft that's already grounded in real quotes is faster than starting from a blank page.

Variants

• Add a helpdesk. If support runs through a shared Gmail inbox, it's already covered. For other helpdesks, connect them as they become available or bring them via an MCP server.
• Per-segment syntheses. Enterprise versus self-serve feedback as two workflows with different Slack channels and labels.
• Route bugs. Add a structured-output step that classifies each theme as bug / capability / confusion, then a conditional that files bugs to Linear (a task.write you'd grant narrowly) and leaves the rest for you.

Common questions

Does the agent post anything to Slack?

Not in this recipe. It reads Slack and writes to your drive and to a chat thread with you. Posting would be a gated message.send action.

How does it handle the same customer saying the same thing twice?

It counts distinct customers per theme, using sender and thread identity from each source. Tell it your customer naming conventions if sources disagree.

Can it read call transcripts from tools other than Fireflies?

Any transcript source that's connected. Fireflies is what ships today; ask us about yours.

What about privacy?

The agent only reads sources you connected and granted, every read is logged in the run, and the model provider doesn't retain the data (zero-data-retention endpoints for Anthropic, OpenAI and Google models).