Experiment

Meeting Capture

Tools Recede, Understanding Remains

December 2025 · Swift + Cloudflare Workers · Heidegger, Rams

╔══════════════════════════════════════════════════════════════════╗
║  MEETING CAPTURE                                                 ║
║                                                                  ║
║  ┌──────────┐     ┌──────────┐     ┌──────────┐     ┌─────────┐  ║
║  │  ((o))   │     │    ~     │     │   ___    │     │   [D1]  │  ║
║  │  AUDIO   │ ──► │ WHISPER  │ ──► │  TEXT    │ ──► │  STORE  │  ║
║  └──────────┘     └──────────┘     └──────────┘     └─────────┘  ║
║                                                                  ║
║     Swift          Workers AI        Transcript       Database   ║
║    (local)        (Cloudflare)                                   ║
╚══════════════════════════════════════════════════════════════════╝

Hypothesis

A meeting transcription tool with fewer features will achieve higher actual utility than feature-rich alternatives—because tools that recede into transparent use (Zuhandenheit) serve understanding better than tools that demand attention.

Success Criteria

✓
Zero interactions during meeting

Tool requires no attention while recording

✓
Transcription completes <60s post-meeting

Actual: 15-30s for typical meetings

✓
Auto-detection works for Zoom, Meet, Teams

Tested with Zoom; Meet/Teams detection implemented

△
Auto-stop when meeting ends

Partial: Manual stop required; detection unreliable

✗
System audio capture

macOS Sonoma requires kernel extension; microphone only

Measured Results

MetricTargetActualStatus
Setup time<5 min~3 minPass
Transcription latency<60s15-30sPass
Audio file size (30min)<10MB~500KBPass
Interactions during meeting00Pass
Post-meeting interactions1 (stop)1Partial
System audio captureYesNoFail

Grounding: Heidegger on Tools

In Being and Time (1927), Heidegger distinguishes two modes of encountering equipment:

Zuhandenheit

Ready-to-hand

The tool disappears into use. When hammering, we think about the nail—not the hammer. The tool serves the work; attention flows through it.

Target state for Meeting Capture

Vorhandenheit

Present-at-hand

The tool becomes an object of attention. When the hammer breaks, we suddenly see the hammer. It intrudes into consciousness.

What feature-bloat creates

"The less we just stare at the hammer-thing, and the more we seize hold of it and use it, the more primordial does our relationship to it become."

— Heidegger, Being and Time §15

Feature-rich tools like Granola and Otter shift equipment from Zuhanden to Vorhanden through accumulated complexity: speaker identification, action item extraction, calendar sync, team sharing. Each individually justified; collectively attention-demanding.

Method: Subtractive Triad

Following Rams's tenth principle ("As little design as possible") and the Subtractive Triad, we asked three questions for each potential feature:

DRY (Implementation)

"Have I built this before?"

→ Cloudflare D1, R2, Workers AI already exist. Use them.

Rams (Artifact)

"Does this feature earn its existence?"

→ Speaker diarization? Action items? No—complexity without proportional value.

Heidegger (System)

"Does this serve the whole?"

→ Store in D1 to join the CREATE SOMETHING knowledge graph.

Features Removed

  • Real-time transcription — Complexity without benefit; post-meeting suffices
  • Speaker identification — Adds 3x complexity; rarely needed for personal use
  • Action item extraction — AI hallucination risk exceeds value
  • Calendar integration — Meeting detection handles this
  • Team sharing — Personal tool; single user
  • Web dashboard — curl suffices for queries

Features Retained

  • Auto-detect meetings — Zoom, Google Meet, Teams via NSWorkspace + AppleScript
  • Record microphone — AVFoundation; system audio requires kernel extension
  • Upload on stop — Multipart form to Cloudflare Worker
  • Transcribe via Whisper — @cf/openai/whisper-large-v3-turbo
  • Store in D1 — Searchable via SQL

Implementation

macOS Menubar App

Swift + SwiftUI · ~400 LOC

  • • NSWorkspace for app monitoring
  • • AppleScript → System Events for call detection
  • • AVFoundation for microphone capture
  • • URLSession for multipart upload

Cloudflare Worker

TypeScript · ~200 LOC

  • • R2 for audio storage
  • • Workers AI (Whisper) for transcription
  • • D1 for transcript storage
  • • ctx.waitUntil() for background processing

Technical Learnings

Whisper requires Base64, not ArrayBuffer

@cf/openai/whisper-large-v3-turbo expects Base64-encoded audio. The standard Whisper model accepts integer arrays. This distinction caused "Type mismatch of '/audio'" errors until resolved. Documentation unclear.

Unsigned apps cannot run AppleScript

macOS blocks unsigned apps from querying System Events, even with Automation permission. Ad-hoc signing with com.apple.security.automation.apple-events entitlement resolves this without Developer ID.

ArrayBuffer detaches in async context

Audio data passed to ctx.waitUntil() becomes unusable. Solution: store in R2 first, re-fetch within the async handler.

Limitations & Failures

System audio capture not implemented

macOS Sonoma requires a kernel extension or audio driver (like BlackHole) for system audio capture. This experiment captures microphone only—participant voices are clear, but shared content audio is not captured. A production tool would need to integrate with an audio routing solution.

Auto-stop detection unreliable

Zoom's helper processes (CptHost, caphost) don't terminate predictably when a meeting ends. AppleScript window-count detection has false positives. Current implementation requires manual stop. Acceptable for personal use; would need improvement for broader adoption.

Transcription accuracy varies

Whisper performs well for clear speech but struggles with overlapping voices, technical jargon, and non-English content. No quantitative WER measurement performed in this experiment.

Reproducibility

Prerequisites

  • macOS 14+ (Sonoma) with Xcode 15+
  • Cloudflare account with Workers AI access
  • Microphone permission granted to app
  • Automation permission for System Events

Expected Challenges

  • Code signing: Without ad-hoc signing, AppleScript will fail silently
  • Whisper input format: Use Base64, not ArrayBuffer for turbo model
  • D1 FTS5: Full-text search virtual tables may cause database corruption

Starting Prompt

Build a macOS menubar app that:
1. Detects when Zoom/Meet/Teams starts a call
2. Records microphone audio
3. Uploads to a Cloudflare Worker on stop
4. Transcribes via Whisper and stores in D1

Use: Swift/SwiftUI, AVFoundation, NSWorkspace
Constraints: No UI during recording, <500 LOC total

Conclusion

The hypothesis is partially validated. A minimal transcription tool achieves Zuhandenheit for its core function—zero attention during meetings, automatic transcription after. The tool recedes.

However, the manual stop requirement and microphone-only capture represent failures to fully disappear. A tool that demands even one interaction per meeting has not completely achieved ready-to-hand status.

"Weniger, aber besser."

— Dieter Rams

Less, but better. The discipline of removal produced a functional tool in ~600 LOC across two components. Whether "better" depends on use case: for personal meeting notes, this suffices. For team collaboration, the removed features become necessary.