Our thinking is our IP, our commercial advantage. But as transcript-based apps are on the rise, more often than not, it’s locked away inside an increasingly disparate tech stack. I wanted to know what happened if I actually went back to look at over a year’s worth of ‘thinking’.
The problem with rented land
I dictate everything. I’ve been using Wispr Flow for 14 months now and it runs on my Mac, my laptop and my phone. While I still sometimes type (hello shared office space), a lot of the prompts I give Claude or ChatGPT, messages I draft, and every half-formed idea muttered into my phone while walking the dog, are only recorded in Wispr. Fourteen months of that, for me, came to roughly 1.7 million words.
Now, in the olden days - if I was thinking through a problem, if I was in a client meeting, if I was carrying out a piece of work - none of that was recorded as a transcript for later retrieval. So all of this process comes with the caveat that this is an AI-world-specific issue. But the fact is that AI already competes with some knowledge work; I believe that how I think and how I tackle problems is going to become increasingly valuable. What I explain below is what it looks like to start treating your own thinking as an asset class that (theoretically at least) could be commoditised or licenced in the future.
But back to the problem at hand: 1.7 million words (some noise, but some real ‘insight’), and all of it was locked inside of someone else’s app. The raw material existed, but as I was about to find out… it wasn’t set up for me to use it after that initial moment of input. I want to acknowledge here that Wispr specifically records and stores audio files and transcripts locally (i.e. on the user’s own device) unless a user specifically turns on cloud sync. Brilliant for privacy; rubbish if you’re hoping to run a clean data export from multiple devices. Because most Wispr users don’t want to bulk export historical voice notes (yet!), the apps don’t have an in-built way of doing this.
Getting the data out
The desktop apps were actually relatively straightforward. Wispr Flow keeps a local database on each machine, so the full dictation history was already sitting there in a file I could access via Terminal (thanks Claude Code). It required a little technical know-how, but nothing a keen bean with a Claude or Codex subscription couldn’t also figure out.
The phone was the awkward one. There is no export feature in the iOS app and because I didn’t have cloud sync toggled on, Wispr had no way to directly access my audio files or transcripts (props to their support team though, who have been awesome in trying to problem solve this).
The answer turned out to be a plain iPhone backup. I ended up backing my phone up to my Mac, saving the backup locally rather than on iCloud. That meant that the app’s local database was also saved on my laptop’s harddrive, although it still wasn’t exposed directly. But after giving Terminal full disk access and downloading iWizard, I was able to access and then export all 2238 original voice recordings and their corresponding transcripts. No jailbreak, and no paid software beyond a free backup browser. Everything you have ever dictated into your phone is recoverable this way.
Nerding out: turning transcripts into a database
Raw transcripts are unusable at this volume. Reading 1.7 million words back is not a research method, and wouldn’t be worth the time given the level of noise captured amongst the real, interesting signals.
Fable 5.1 enters the chat
Using Claude Code in /workflow mode, I ran a basic data exploration - data analysis process, stripping out the noise and identifying patterns, trends and specific data about my business. This resulted in over 2000 ‘findings’, each a summary-verbatim quote pairing plus metadata (device, date, that sort of thing). It included sales insights, pricing decisions said out loud, client objections, one-liners worth reusing, product ideas, and a category I called “what I keep saying I’ll do and don’t”, which I would recommend to anyone attempting this.
Three rules made the difference between a database I trust and an AI summary I would have ignored.
Every quote is verified. A script checks each one is a character-for-character match against the source transcript which prevents paraphrasing and invented quotes.
Every finding carries a date. That turns a pile of insights into a timeline, so I can see when a pattern started and whether it has got better or worse since.
Attribution lives in the data rather than in a prompt. Quotes from or about other people are flagged as needing permission before any external use, and the flag is written into the quote text itself so that it survives every copy and paste or retrieval call. My AI cannot accidentally publish someone else’s words as mine.
What it found (brutal, but brilliant)
The database turns out to be very good at surfacing the gap between what I advise and what I do, which was less enjoyable to read than it sounds.
I tell clients to stop selling time for money. My revenue is mostly day rates.
I tell clients to hold their prices, and the transcripts caught me discounting in real time, on calls, within minutes of having decided not to.
Across fourteen months I had diagnosed my own biggest bottleneck out loud, to an AI, at least eleven separate times… and acted on it roughly never.
None of that was visible to me until the words came back out of the chat windows.
Then the analysis surfaced something that made me sit back from my desk. Months earlier I’d recorded a voice note in which I had sketched a product idea: a vault that holds everything valuable pulled out of your own voice transcripts. I had designed the thing we were in the middle of building, into my phone, and practically forgotten it existed.
Making my thought history searchable
A database you have to read is only half useful. Especially with 2000+ rows. Once I had the data filed somewhere, the last step was making it answer questions. I decided to do this in Xano.
Every finding and every transcript passage was uploaded to a database in Xano, and then converted into an embedding (which is a set of numbers that captures meaning rather than keywords, and is how LLMs work). That took one evening and cost me six pence in total (through OpenAI’s embeddings). One of the things I love about Xano is how easy it is to create a MCP service (a standard way of plugging a data source into an AI assistant) from a database, so I turned my database into an MCP.
Et voila, all the useful ideas and data from the last 14 months of Wispr voicenotes now connects directly to Claude via MCP and is now searchable by both keyword and meaning (vector embeddings). I can analyse it and it also functions now as a decent RAG (retrieval augmented generation) AI workflow for any AI tools I connect to it.
For example: I might never have explicitly said: “I feel like I should be charging more” so a simple keyword search would return an empty response. But by turning this into vector embeddings, within 0.4 seconds Claude was able to pull some of my thinking on pricing structures and business models.
Now when I draft a proposal or a talk, my AI tools can pull my own words, my go-to examples, whatever is needed, rather than generating plausible filler.
Nerdy too? Steal this
In certain circles, building a second brain is nothing new. What I’m hoping to show here is that the value a lot of us offer (professionally, definitely; generally, maybe) is often already captured in the third party apps we subscribe to - tools like Wispr, meeting recordings, email chains, whatever.
So start by treating this thinking history as an asset rather than an exhaust. It is sitting in local app databases and phone backups right now, and getting it out is an afternoon of work.
Then extract findings instead of summaries. Verbatim quotes, checked against the source, with a date and an owner attached to each one. The discipline matters far more than the tooling here. If the output is a summary rather than something you can trace back to a sentence you actually said, you will not trust it six months from now when it tells you something you would rather argue with.
After that, make it searchable by meaning. Embeddings cost pennies at this scale and turn a dead archive into something you can interrogate in plain language, including with the questions you would be slightly embarrassed to type into a search box.
The tools cost me almost nothing beyond subscriptions I was already paying for. The cost came from having (almost) lost 14 months of thinking, where I never stopped to listen back a second time (because who does, right?!).
Now though: my archive can now tell me I have raised the day-rate question eleven times. It might not be able to make me actually change my prices on the next call. But it is going to keep count either way.