Voice capture

Press Shift+N anywhere in MLW to dictate. Whisper transcribes, Claude reads the transcript and pulls out structured tasks for you to review before they save. Reflective or journal-style content gets appended to today's Daily Page instead.

The whole loop is built around the moments when typing is friction: walking to the car, mid-conversation, in the shower with your phone on speaker, the morning brain dump before the day starts.

How it works

1. Press Shift+N (or tap the floating capture button → Voice on mobile)
2. Talk. Stop talking when you're done.
3. The transcript appears in a review sheet
4. Claude's single-pass extraction runs. Each candidate task lands as its own row with full triage metadata already filled in (status, area, project, tags, energy, due date when you mentioned one). Nothing is left to a second AI pass; what you see in review is what saves.
5. Edit anything, deselect anything, then Save. Selected items become real tasks; the original transcript still lives on as a note for traceability.

If you said "I should remember to email Jane about the contract, and I'm thinking about how to structure the Q3 OKRs," you'll see two task candidates ("Email Jane about the contract", "Structure Q3 OKRs") plus a note containing the full transcript. The Jane email might come pre-tagged with @jane if Claude recognized her from your existing people tags.

Reflective vs structured

Claude classifies each chunk:

• Discrete tasks. Concrete actions ("Email Jane", "Pick up cake Friday") → review sheet.
• Reflective content. First-person thinking, journaling, processing ("I'm worried about the renewal because…") → appended directly to today's Daily Page without review interruption.

The distinction means a stream-of-consciousness brain dump doesn't fill your Inbox with phantom tasks; it lands where journal content belongs.

What to do when transcription is wrong

The review sheet shows the full transcript at the top. If a task title got mangled (rare with Whisper-large but happens with proper nouns or accents), edit it inline. The underlying transcript stays the same in the saved note.

If the whole transcription is garbage (mic muted, very loud background, etc.), close the sheet without saving and try again.

Privacy

The audio is sent to Whisper for transcription and to Claude for extraction. We don't keep the audio file after extraction completes. Transcripts are stored on the resulting note (so you can audit what was captured); delete the note to delete the transcript.

If you've enabled Pro, transcription uses Groq's Whisper-large-v3-turbo (fast); free users use OpenAI Whisper.

Keyboard shortcut

Shortcut	Action
Shift+N	Open voice capture (record → AI parses → review)

Voice capture on iOS PWA

iOS used to re-prompt for mic permission on every voice capture, which made the feature unusable in practice. That's fixed. MLW now primes the mic permission inside the click handler so iOS counts the gesture as user-initiated and remembers your "Allow" answer for the session.

One residual: iOS resets microphone permission after a service-worker update (when MLW deploys a new bundle and the app refreshes). The next voice capture after a deploy re-prompts. Click Allow once and the rest of the session works.

What MLW doesn't do (yet)

• No on-device transcription. Both Whisper paths are cloud. On-device Whisper-tiny in WASM is on the roadmap as part of the Private Mode story; without it, audio leaves your device.
• No real-time transcription. We transcribe after you stop speaking, not as you go. Faster overall (no streaming overhead) but doesn't show partial text mid-sentence.
• No voice commands. Talking doesn't trigger app actions other than "save what I just said." We'd rather not encourage rambling at the app.

Common questions

How do I voice-capture a task in MLW?

Press Shift+N anywhere in the app, or tap the floating capture button on mobile and pick Voice. Talk, stop talking, and Whisper transcribes the audio while Claude extracts structured task candidates for you to review before saving.

What gets saved when I dictate?

Discrete tasks become rows you can edit and select in a review sheet; the original full transcript is saved as a note for traceability. Reflective or journal-style content gets appended to today's Daily Page directly without the review interruption.

How accurate is the transcription?

Voice capture uses Whisper-large-v3-turbo (Pro) or Whisper-1 (free). Both are very accurate for general speech. Proper nouns, technical jargon, and accents are the typical error cases; you can edit any extracted task title before saving.

Does MLW keep my audio recordings?

No. Audio is sent to Whisper for transcription, then dropped after extraction completes. Only the resulting transcript (saved as a note) and extracted tasks persist.

Can I use voice capture on my phone?

Yes. Tap the floating capture button in the bottom-right corner and pick Voice. Same Whisper + Claude pipeline as desktop.

Related

• Quick Capture. Keyboard-driven structured capture, the typing alternative.
• Daily pages. Where reflective voice content lands.
• AI Weekly Review. The same Claude that pulls tasks from your dictation also writes your weekly summary.