Weekly customer feedback digest from interview recordings

Build a weekly customer-feedback digest from raw interview and user-call recordings that my team drops into a Google Drive folder.

Trigger: cron, every Monday at 9am America/New_York. Google Drive is not a poll provider, so cron plus a filtered List Files call is the right pattern here, not a poll trigger.

Step 1: Use the google-drive List Files operation to find every audio file added to a configured Drive folder ID in the last 7 days. Use a Drive query like `'<folderId>' in parents and mimeType contains 'audio/' and createdTime > '<sevenDaysAgoISO>' and trashed = false`. Page through results until all matching files are collected. The folder ID should be a configurable input.

Step 2: For each matching file, call google-drive Download File Content to fetch the raw audio bytes.

Step 3: For each recording, push it through AssemblyAI. Call Upload Audio File to stream the bytes to AssemblyAI and get an upload URL. Then call Submit Transcript with that upload URL and these options enabled: speaker_labels: true, sentiment_analysis: true, entity_detection: true. Then call Get Transcript and poll until status is `completed`. If the polled status is `error`, log it and skip that file rather than failing the whole run.

Step 4: Once all transcripts are in hand, the agent should: (a) cluster the conversations into 3 to 6 recurring themes (for example 'onboarding friction', 'billing confusion', 'feature request: SSO') with punchy, specific theme names; (b) for each theme pick the single sharpest verbatim quote, attribute it with the AssemblyAI speaker label, and count how many distinct interviews touched that theme; (c) call out one or two notable outliers, meaning surprising or strong-signal moments worth flagging even if they did not cluster into a theme.

Step 5: Post the digest as a single message to a configured Slack channel (default #product-feedback) using the slack Send a Message operation. Format it as: a one-line intro with the date range and the number of interviews processed, one bullet per theme with the bolded theme name, the interview count in parentheses, and the verbatim quote on the line underneath, and finally a short 'Outliers' section at the bottom. Use Slack mrkdwn formatting (`*bold*`, not `**bold**`).

If no audio files were found in the last 7 days, post a short 'no new interviews this week' note to the same channel instead of skipping silently, so the team always knows the workflow ran.

This is an agent workflow (not code) because clustering, quote selection, theme naming, and outlier identification are open-ended judgement calls. The agent should reason over the full set of transcripts rather than following a rigid script.