Walking Works
Talk on a walk.
Come home to finished work.
I clip a small microphone to my collar, go for a walk, and think out loud. When I get back, an AI turns the recording into notes, letters, tasks and essays, filed where they belong. This page lists the equipment, the software and the process, so you can build the same thing.
Read this first
I am happy to answer questions about my setup. But the honest truth is that your real next step is to ask an AI to help you set this up, not me. Paste this page, or the prompt at the bottom, into Claude, ChatGPT or whatever you use. It knows the tools better than I do, it will not forget a step, and it will answer in seconds instead of days.
Use me for the "why" and for "does this actually feel good to live with". Use the AI for the "how".
1The idea in one minute
A microphone that records to its own memory, so there is no phone in your hand. A transcription service that turns an hour of speech into text for about a quarter. An AI agent that reads the text and does what you asked in it. A plain folder of markdown notes where everything lands. Two-way sync so the notes are on every machine you own.
The important design choice is that the AI does not just transcribe, it acts. If I say "write me a letter to my landlord about the heating", a draft letter exists when I get home, with my exact words attached underneath so I can check it did not invent anything.
2Equipment
A wireless microphone with onboard recording
I use a DJI Mic 3 kit: two transmitters, one receiver, a charging case.
The transmitter clips to your collar and records to its own internal memory. The receiver stays at home; I never use it for this. It records 32-bit float, so it does not clip when you laugh or shout, and it keeps recording even with no phone or receiver anywhere near.
About 280 dollars for the kit, less for a single transmitter.
A computer running Linux, macOS or Windows
Mine is an eight year old ThinkPad with 8 gigabytes of memory.
It only needs to mount a USB drive, run a few scripts, and run the AI agent. Nothing heavy happens locally; the transcription happens in the cloud.
Whatever you already own.
A USB-C data cable
Data, not charge-only.
Plug the charging case in and each transmitter shows up as its own USB drive. A charge-only cable will silently charge and never mount. Check this first when nothing appears.
A few dollars.
Optional: an always-on second machine
I rent a small cloud server for scheduled jobs and a second copy of the notes.
Not needed for dictation. Skip it until you want things to happen while your laptop is closed.
Around 48 dollars a month. Optional.
Check the model before you buy. The white DJI Mic Mini has no internal storage and cannot work as a recorder at all. The Mic Mini 2S does. The Mic 3 does. If a review does not say "onboard recording" or "internal storage", assume it has none.
3Software
| Job | What I use | Why |
|---|---|---|
| Pull the files off the mic | A small Python script on a two-minute timer | Detects the mounted case, copies new recordings into a dated folder, verifies every byte, then ejects the case. No clicking. |
| Speech to text | Deepgram, model nova-3, with speaker separation and paragraphs on | Fast, cheap, good with outdoor audio. One hour of speech costs about a quarter. |
| Audio clean-up | ffmpeg | A high-pass filter and speech normalization before upload rescue quiet stretches. Compressing to Opus makes uploads 50 times smaller with no loss you can hear. |
| Do the work | Claude Code, a terminal AI agent, with a written "skill" file describing my rules | It reads the transcripts, writes the deliverables, files logs and tasks, and asks me only when unsure. |
| Where notes live | Obsidian, which is just a folder of markdown files | Plain text. Any tool can read it, including the AI. No lock-in. |
| Sync | Syncthing | Free, peer to peer, two-way. Notes appear on every machine within seconds. |
| Cold storage | A script that moves audio older than a week to Google Drive | The audio archive passed 96 gigabytes in six months. Text stays local forever, audio does not. |
None of these are the only option. The AI you ask for help will suggest equivalents, and the setup is the same shape whatever you pick: record, copy, transcribe, act, file, sync.
4The process, start to finish
- Clip on, press record, walk.
The red light means it is recording. I talk about whatever is on my mind: observations, things to do, instructions, sometimes a whole essay.
- Say instructions in a way that survives transcription.
See the next section. The minimum is one spoken word: "instruction".
- Plug the case into the laptop.
Each transmitter mounts as its own drive.
- The automatic half runs by itself. automatic
The timer notices the drive, copies and verifies the new files, renames the originals on the device to
COPIED_so they stay as a backup, ejects the case, and sends everything to Deepgram. One desktop notification when it is done. Silence if there was nothing new. - I tell the AI agent "do the pull process".
It reads the pending list and gets to work.
- The agent first fixes the clock.
The microphone has no idea what time zone it is in. Every pull, the agent works out the real time from what I said, and stores both the raw device time and the corrected local time.
- Transcripts go into the notes folder.
One markdown file per recording, verbatim, with timestamps and speaker labels. Audio never goes into the notes folder; it is too big.
- The agent does what the recording asked.
Letters, essays and drafts go to a "Writing in Progress" folder. Small asides go to today's daily note. "Add a task" becomes a task dated today.
- Every deliverable carries its source.
A collapsed note at the top links to the transcript and quotes what I actually said. The full verbatim dictation sits at the bottom. So I can always check the machine against my own words.
- Nothing leaves the machine without permission.
The agent may draft emails, never send them. It may write files, never upload them. If I want something sent, I have to name the destination out loud in the recording, and it still confirms.
5How to talk so the machine understands
Transcription loses tone of voice, and a rambling walk mixes musing with real requests. A few spoken markers make the difference.
- Other useful markers: note, research, draft, flag, priority, and scratch that to cancel the last thing.
- Do not use the assistant's name as filler. Say it only when you mean it.
- One instruction at a time. Do not batch five requests into one breath.
- Do not edit in mid-air. Say "scratch that" and start the sentence again.
You are recording to think, not to brief. If the protocol starts to feel constraining enough that it dampens the walk's thinking quality, drop it. A jumbled brilliant transcript is better than a clean boring one.
6What it costs
| Item | Cost |
|---|---|
| DJI Mic 3 kit, one time | about $280 |
| Transcription, per hour of speech | about $0.26 |
| A typical 30 minute walk | about $0.13 |
| A heavy month, 50 hours recorded | about $13 |
| Obsidian, Syncthing, ffmpeg | free |
| AI agent subscription | whatever plan you already pay for |
| Optional cloud server | about $48 a month, skip at first |
7Pitfalls I learned the hard way
It only syncs through the phone app, and the sync often does not stick. Mine sat nine hours off for months after a move. Never trust file timestamps; work out the time from what was said, and store both values.
Wind, a mic far from your mouth, or a mumbled stretch can return a near-empty transcript while the audio is fine. In one sweep 48 recordings had recoverable gaps, and one pull lost about 5,000 words. Normalize the audio before upload and have the agent check every transcript for suspicious gaps.
Speaker separation tells you there were two voices, not who they were. I was "Speaker 0" in most recordings one day and "Speaker 1" in three others. Anything that assumes otherwise swaps people's words.
"I flew from Frankfurt" once came through as "a flu from Frankfurt", and an evening of advice was built on an illness that never existed. Before acting on a surprising new fact from dictation, check whether anything else supports it.
A low case battery trips the laptop's USB over-current protection and nothing appears. Charge the case from the wall first.
Sync tools cannot merge text. Give each automated writer one owner machine, and run a weekly report that lists conflicts but never deletes them.
Byte-check every copy. Keep the originals on the device until you are sure. Move audio to cold storage only after the remote copy is confirmed the same size.
8Is it actually good for you?
I am not sure, and I would rather say so. My own ranking after months of this: walking in silence beats dictating, and dictating beats staring at a phone. The habit is clearly useful for capture. Whether it is good for my mood is ambiguous. Some days recounting my life into a collar clip feels like avoiding the walk itself.
What I can say is that it moved a lot of thinking out of my head and into finished things, and that the "AI acts on it" part is what makes it worth the fuss. Transcripts alone would just be a bigger pile.
9A prompt to start with
Copy this into Claude, ChatGPT or another assistant. Change the bits in brackets. Then keep asking it questions until it works.
I want to build a "dictate on a walk, AI does the rest" setup. Here is the shape: 1. A wireless lavalier mic that records to its own internal memory (I have / am considering: [DJI Mic 3 or other]). 2. When I plug it into my [Linux / macOS / Windows] computer, a script should automatically copy new recordings into a dated folder, verify them, and eject the device. 3. Each recording should be transcribed with a speech-to-text API (Deepgram nova-3 or similar), with speaker separation and paragraphs, after normalizing the audio with ffmpeg. 4. An AI agent (Claude Code or similar) should then read the transcripts, act on spoken instructions (marked by the word "instruction"), write deliverables into my notes folder, and attach my verbatim words to everything it writes. 5. My notes live in a folder of markdown files (Obsidian) synced with Syncthing. 6. The agent must never send or upload anything without explicit permission. Please walk me through it step by step, starting with what to buy, and stop after each step so I can confirm it works before we continue. Warn me about: mic clock drift, quiet speech being dropped by transcription, speaker labels not being stable, and charge-only USB cables.