Thinking Notebook · For nerds · Roger Basler de Roca

The Nerd Guide

This is how the thinking notebook runs fully automatically: a Mac records a stream, transcribes locally, stores the texts in Dropbox, and scheduled tasks in Claude turn them into takeaways, news and a Notion sync. If the event itself offers transcripts or summaries, for example in an event app, that is the best source: Claude fetches them via your browser, where you are logged in, and the recorder only runs as a backup. In the end, the human decides what stays. For meetings: BlackHole only records what comes out of the speakers, your own voice is missing. For meetings, better use the transcript from Zoom, Teams or tl;dv.

Everything on this page as one file: commands, script, five prompts, checklist, troubleshooting.

← Back to the starter

Before you record

Only record what you are allowed to record: your own meetings with the consent of everyone involved, or streams where the organiser agrees. Many events offer recordings or transcripts anyway, ask there first. Never hand your login details to an AI. You log in yourself in the browser.

The architecture

Route A, if the organizer offers transcripts:

  1. 1Event app with transcripts
  2. 2Claude reads via your browser (Chrome or Brave with the Claude extension)
  3. 3Claude: takeaways, notebook, Notion

Route B, your own recording as a backup:

  1. 1Stream in the browser
  2. 2Sound via BlackHole
  3. 3Recording with ffmpeg
  4. 4Local transcription with whisper.cpp
  5. 5Text file in Dropbox
  6. 6Claude: takeaways, notebook, Notion

The audio stays on your computer. Only the texts go to the cloud.

What you need

1 · Install the tools

Install Homebrew (asks once for your Mac password):

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

On Apple Silicon Macs, then run the two lines Homebrew shows at the end. Then the tools:

brew install ffmpeg whisper-cpp && brew install --cask blackhole-2ch

Then restart the Mac once.

2 · Route the sound

  1. 1Open the Audio MIDI Setup app
  2. 2Click + at the bottom left and create a Multi-Output Device
  3. 3Tick your speakers or headphones and BlackHole 2ch
  4. 4System Settings, Sound, Output: choose this new device

This way you keep hearing the stream, and at the same time the sound reaches the recording.

As long as the multi-output device is selected, the volume keys do not work. Adjust the volume directly in the stream player.

3 · Load the speech model

small is a good compromise for older Macs, on Apple Silicon large-v3-turbo works too.

mkdir -p ~/whisper-models && curl -L -o ~/whisper-models/ggml-small.bin https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.bin

On Apple Silicon, also load large-v3-turbo and start with MODEL=~/whisper-models/ggml-large-v3-turbo.bin:

curl -L -o ~/whisper-models/ggml-large-v3-turbo.bin https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo.bin

4 · The recording script

Save the script in your home folder (~/stream_recorder.sh). It records in 10-minute chunks, transcribes each chunk, only stores chunks with real text and keeps the Mac awake. Markers like [MUSIC] or [Applause], which whisper writes during silence, do not count as words.

On first start, macOS asks whether Terminal may access the microphone. Allow it: for macOS, BlackHole counts as a microphone. If you decline, you only record silence.

#!/bin/bash
# Stream recorder for the thinking notebook
# Records audio in 10-minute chunks and transcribes them locally with whisper.cpp.
# Transcripts land in the Dropbox folder and are picked up by a scheduled task in Claude.
# Start:  bash ~/stream_recorder.sh        Stop:  Ctrl+C

set -u
if [ -z "${OUT:-}" ]; then                              # find target folder (Dropbox) automatically
  DB=$(ls -d "$HOME/Library/CloudStorage/Dropbox"* 2>/dev/null | head -1); [ -z "$DB" ] && DB="$HOME/Dropbox"
  OUT="$DB/Thinking-Notebook-Transcripts"
fi
WORK="${WORK:-$HOME/Thinking-Notebook-Recording}"     # audio stays local, never goes to the cloud
DEVICE="${DEVICE:-BlackHole 2ch}"                     # audio input that receives the stream sound
MODEL="${MODEL:-$HOME/whisper-models/ggml-small.bin}"
LANG_CODE="${LANG_CODE:-en}"
SEGMENT="${SEGMENT:-600}"                             # seconds per chunk

mkdir -p "$OUT" "$WORK/chunks" "$WORK/done"
command -v ffmpeg >/dev/null || { echo "ffmpeg missing: brew install ffmpeg"; exit 1; }
command -v whisper-cli >/dev/null || { echo "whisper-cli missing: brew install whisper-cpp"; exit 1; }
[ -f "$MODEL" ] || { echo "Model missing: $MODEL"; exit 1; }

echo "Recording from '$DEVICE' in chunks of $SEGMENT s. Transcripts to: $OUT"
caffeinate -dimsu -w $$ &

( while true; do
    ffmpeg -hide_banner -loglevel error -f avfoundation -i ":$DEVICE" -ac 1 -ar 16000 \
      -f segment -segment_time "$SEGMENT" -reset_timestamps 1 -strftime 1 \
      "$WORK/chunks/%Y-%m-%d_%H-%M-%S.wav"
    echo "$(date '+%H:%M:%S') recording interrupted, restarting in 5 s"; sleep 5
  done ) &
REC=$!
trap 'echo; echo "Stopping ..."; kill $REC 2>/dev/null; pkill -P $REC 2>/dev/null; exit 0' INT TERM

while true; do
  FILES=( $(ls -1 "$WORK/chunks/"*.wav 2>/dev/null | sort) )
  N=${#FILES[@]}
  if [ "$N" -gt 1 ]; then
    for ((i=0; i<N-1; i++)); do
      F="${FILES[$i]}"; B=$(basename "$F" .wav); DAY=${B%%_*}
      mkdir -p "$OUT/$DAY"
      whisper-cli -m "$MODEL" -l "$LANG_CODE" -nt -otxt -of "$WORK/done/$B" -f "$F" >/dev/null 2>&1
      WORDS=0
      if [ -s "$WORK/done/$B.txt" ]; then
        # Markers like [MUSIC], (music) or *applause* do not count as words
        WORDS=$(sed -E 's/\[[^]]*\]//g; s/\([^)]*\)//g; s/\*[^*]*\*//g' "$WORK/done/$B.txt" | wc -w | tr -d ' ')
      fi
      if [ "$WORDS" -gt 15 ]; then
        { echo "Recording $B (automatically transcribed, not checked)"; echo; cat "$WORK/done/$B.txt"; } > "$OUT/$DAY/$B.txt"
        echo "$(date '+%H:%M:%S') transcribed: $B ($WORDS words)"
      else
        echo "$(date '+%H:%M:%S') silence, music or too short, skipped: $B"
      fi
      mv "$F" "$WORK/done/"
    done
  fi
  sleep 20
done

Quick test with any video, chunks of 60 seconds:

SEGMENT=60 bash ~/stream_recorder.sh

After about two minutes "transcribed" appears and a text file is in the folder. Stop with Ctrl+C. For continuous mode, start without SEGMENT and leave the terminal open:

bash ~/stream_recorder.sh

For talks in several languages:

LANG_CODE=auto bash ~/stream_recorder.sh

After the event, stop with Ctrl+C and delete the audio folder ~/Thinking-Notebook-Recording.

5 · Claude takes over

Four scheduled tasks in Claude turn raw text into work, plus an optional fallback. If the organizer provides transcripts, you do not need the stream start or the fallback. Create them by giving Claude the respective prompt and adding: "Create this as a scheduled task, hourly from 8 am to 8 pm" (news: daily at 6:30). Important: turn on "Auto-approve" in the settings of each task, otherwise every run stops and waits for your confirmation. Every task starts in a fresh session without memory. That is why the link to your notebook is in every prompt.

Feed in transcripts

hourly

My thinking notebook: [link to your notebook]. First read which collections it stores its data in, and write in the same format. Route A: in my browser with the Claude extension, open the programme page of the event app [URL]. I am logged in there. If a login page appears, enter nothing, switch to route B and report "Login expired". For all of today's finished programme items, read the transcript or summary that is not yet in the notebook. Route B, only if A does not work: read all .txt files in the day folders of the Dropbox folder /Thinking-Notebook-Transcripts from [start date, e.g. 2026-09-25] that are not yet in the "processed" list. Files with only markers like [MUSIC] or fewer than 30 real words, you only add to the list; combine the rest into blocks of at most 60 minutes. For each talk or block extract takeaways: figure, quote (verbatim only), insight, task (doable in under 30 minutes), open question, AI agent shown (level 1 reads, 2 drafts, 3 acts after approval, 4 acts autonomously). Store each talk with time, title, speaker and a one-sentence conclusion as an input, together with its takeaways. Invent nothing, never post or rate anything on the platform.

Morning news

daily 6:30

My thinking notebook: [link to your notebook]. Research 6 current news items from the last 72 hours on my topics: [your topics]. Every item with a real URL, source and a one-sentence summary. Replace the news in the notebook with them.

Notion sync

hourly · one page per day

My thinking notebook: [link to your notebook]. Under my Notion page [page name], create a subpage "The day in takeaways · [weekday], [DD.MM.YYYY]" for today if it does not exist yet, and replace only its content. Use Notion's design features: at the top a grey callout with the date and last update, below it three coloured tiles side by side (talks or meetings, takeaways, decided and open). Think ahead (blue): the day's appointments from my calendar, only time and title, and the news, both collapsible. Think along (purple): a yellow callout with the five strongest takeaways, then per talk or meeting a collapsible block with time, title, speaker, a one-sentence conclusion and the takeaways with symbols (📊 figure, 💬 quote, 💡 insight, 🤖 AI agent, ✅ task, ❓ question). Kept ones in bold, discarded ones struck through with reason. Think back (green): three thought prompts as callouts, each with a thesis, evidence from at least two of the day's sources and a one-sentence consequence, clearly marked as a suggestion; below them three empty lines for my own thoughts. Only entries from today, never overwrite pages of other days, do not rephrase takeaway texts. Additionally: create a day sheet as HTML in the notebook's look and embed it at the top of the page. Only create the day sheet on the last run of the day.

Start the stream in the morning (optional)

daily 8:50 · needs Chrome or Brave with the Claude extension and the Mac linked to Claude

On my Mac, open the stream page [URL] in my browser with the Claude extension (Chrome or Brave). I am logged in there; if a login page appears, abort and report it. Start playback with sound, wait for the programme to start if necessary. Leave the tab open. Then check whether new files arrive in the transcript folder and tell me in three sentences whether everything is running. Create this task in a chat that is linked to your Mac, so it is bound to the Mac. Do not play it in the Claude app's built-in browser, its sound did not reach BlackHole in our test.

Fallback directly from the Mac (optional)

hourly at :45 · needs the Mac linked to Claude

Do the same as the task "Feed in transcripts", but read the files directly on my Mac in the folder ~/Library/CloudStorage/Dropbox/Thinking-Notebook-Transcripts instead of via the Dropbox cloud. Use the same "processed" list in the notebook, so nothing arrives twice. Report "Check recorder" if no file with real text has been created for an hour during stream times.

6 · Checklist

When something gets stuck

Silence or too short
The sound is not reaching BlackHole. Check the output and whether the tab is muted.
Device not found
ffmpeg -f avfoundation -list_devices true -i "" then start with DEVICE="Name" bash ~/stream_recorder.sh
Mac too slow
Load the smaller model ggml-base.bin and start with MODEL=~/whisper-models/ggml-base.bin.
Mac falls asleep
Keep the lid open, power adapter on, energy saving off. With the lid closed, even caffeinate does not help.
Wrong names in the text
Raw transcripts are not checked. So only quote after checking yourself.
Only silence although sound is playing
Terminal has no microphone permission. System Settings, Privacy & Security, Microphone, turn on Terminal, restart the script.
Empty files with [MUSIC]
whisper writes markers during silence. The current script version no longer counts them. Older version? Reload the script.
Folder missing on dropbox.com
Check the Dropbox icon in the menu bar: app running, signed in, sync not paused. Until it works, the fallback fetches the texts directly from the Mac.
Takeaways from test recordings in the notebook
Set the start date in the feed-in task.
Talks in other languages sound bumpy
Start with LANG_CODE=auto.
Tasks do not run through
Turn on "Auto-approve" in the task settings.
Disk full
Empty the folder ~/Thinking-Notebook-Recording/done, it only contains processed audio.
Stream is playing, recording stays silent
Is the stream playing in the Claude app's built-in browser? Its sound did not reach BlackHole in our test. Play it in Chrome or Brave.
Stream start aborts after a few seconds
The browser extension is not connected. Keep the browser open, the extension signed in, the Mac awake.
Login to the event app expired
Claude never enters passwords. Log in again yourself, the next run catches up on everything.

Why the effort?

Because here the machine does exactly what it is good at: taking notes, sorting, suggesting. And because you keep exactly what only you can do: decide what matters.

Not everything that can be automated has to be dehumanised. You decide.