Build from a meeting

This guide walks you through the full meeting-to-PR flow with a real example.

What you need

• An audio or video recording of a meeting (.mp3, .mp4, .wav, .m4a, or similar)
• A repo where Cascade can work
• Cascade installed and configured (see getting started)

Step 1: Ingest

cascade ingest recordings/sprint-planning-2026-09-25.mp3

Cascade transcribes the audio with Whisper. If you have pyannote.audio installed and a HuggingFace token, it also identifies speakers. Output:

  transcribing recordings/sprint-planning-2026-09-25.mp3  (backend=faster-whisper, model=base)
  transcribed 47 turns (3 speakers, 1825.4s) -> transcripts/meeting-sprint-planning-2026-09-25-103045.yaml

The transcript is YAML you can read, edit, or version-control.

Step 2: Extract

cascade extract transcripts/meeting-sprint-planning-2026-09-25-103045.yaml

The extractor reads the transcript plus your team memory, then asks the LLM to produce structured user stories. Output:

  extracting from meeting-sprint-planning-2026-09-25-103045 (47 turns, 4 memory files)
  extracted 7 stories -> stories/meeting-sprint-planning-2026-09-25-103045.yaml
  tokens: in=8234 out=2156

Open the stories YAML to see what was extracted. Each story has:

• Title and description
• 1+ Given/When/Then acceptance criteria
• Size estimate (XS, S, M, L, XL)
• Confidence score (0-100)
• The transcript turn indices that informed the story
• Optional notes from the extractor

Step 3: Review

cascade review stories/meeting-sprint-planning-2026-09-25-103045.yaml

You’ll see each story one at a time:

------------------------------------------------------------
Story 1 of 7  [size S, confidence 92]
------------------------------------------------------------
Title:       Add 30-second undo for record deletes

Description:
  As a user, I want to undo a record deletion within 30 seconds
  so that I can recover from accidental deletions.

Acceptance criteria:
  1. Given a user just deleted a single record, when they click
     the Undo button in the toast within 30s, then the record is
     restored and the toast disappears.
  2. Given a user just deleted a single record, when 30 seconds
     pass without clicking Undo, then the deletion is permanent
     and the toast disappears.

[a/e/r/s/q] >

Pick one:

• a accept the story as-is
• e edit it in your $EDITOR, then save and exit
• r reject it (won’t be built)
• s skip for now (review later)
• q quit (saves progress; you can resume later)

Decisions are persisted to disk after every action, so it’s safe to quit any time.

Step 4: Build

cascade build stories/meeting-sprint-planning-2026-09-25-103045.yaml

For each approved story, Cascade:

1. Generates an implementation plan (which files, what intent, what risks)
2. Generates the code and tests
3. Creates a new branch
4. Applies the changes
5. Installs deps (best effort)
6. Runs tests
7. Commits
8. Pushes the branch
9. Opens a PR

Output:

  language: Python  (4 memory files loaded)
  stories to build: 3

==> [story-meeting-sprint-planning-001] Add 30-second undo for record deletes
  branch:  cascade/story-meeting-sprint-planning-001/add-30-second-undo
  commit:  a3f8b2c1...
  install: ok
  tests:   passed (53 passed in 0.21s)
  PR:      #142  https://github.com/myorg/myrepo/pull/142

==> [story-meeting-sprint-planning-003] Add export to CSV...
  ...

Or to build just one story:

cascade build stories/meeting-sprint-planning-2026-09-25-103045.yaml --story 1

Step 5: Review the PR

Cascade never merges. A human always approves. Open each PR, read the diff, request changes if needed.

The PR description includes:

• The original story title and description
• All acceptance criteria
• Every file that changed and why
• The test results
• A link back to Cascade

Tips

• Recording quality matters. Clear audio with one speaker at a time produces better stories. Background noise and crosstalk hurts transcription quality.
• Encourage explicit decisions in meetings. “Let’s do X” is much easier for Cascade to extract than “We could maybe think about X.”
• Maintain your team memory. The single highest-leverage thing you can do to improve output quality.
• Start with low-confidence stories. If a story scored 50, it’s probably ambiguous; edit it before approving.