my workflow for AI-assisted information consumption
In this post, I describe a workflow that uses a variety of AI tools to efficiently and intentionally consume information.
First, I use the Wispr Flow mobile app1 to record a rambly voice note for a couple minutes explaining something I’m curious about and listing all the questions that I have about it. Wispr Flow uses a language model to wrangle the transcription into a readable format (for example, it removes “um”s and “uh”s). It does a good job of preserving my voice. In fact, I used it to dictate the first draft of this post on my walk back from the gym today.
I then send the enhanced transcription to Claude, Gemini, ChatGPT, or Perplexity with the “deep research” feature enabled. The model studiously chugs away for a few minutes, synthesizing hundreds of related webpages and papers related to my questions, and notifies me when its report is ready.
Some things I’ve commissioned deep research reports about recently include:
What the economics literature says about career planning and how I can use that to make better decisions with regard to my experience and goals.
The science of foot ergonomics and whether unconventional footwear like wide toe box or minimalist shoes actually have any health benefits.
How to apply technical frameworks from the literature on normative uncertainty (i.e. how to make decisions when you have uncertainty about what is ultimately valuable) to the specific value of meaning in life.
Deep research tools are no substitute for a human research assistant2, but I’ve been impressed with the breadth and relevance of sources they find and how enjoyable the final reports are to read. I’ve been particularly impressed by Gemini’s reports: it will often draw non-trivial connections between the sources and give its own conclusion with my query in mind. I think this is a warning shot for the kind of high-level insights that future AI research assistants will be capable of.
I either read the report directly in the Gemini app or export it to a Google Doc and save the link in my read-later app Readwise Reader. This latter step has a few benefits:
It’s saved in my library alongside all the videos, books, and articles I’ve saved, which are searchable, organized by content type, and synced between my devices. If I don’t want to read it immediately, I’ll just see it later when I’m in the mood to read something (usually before bed on my tablet).
I can make highlights and add annotations. Readwise syncs these to my note-taking app (RemNote), but in practice, I never do anything with these highlights.
ReadWise Reader has a text-to-speech feature that uses models from ElevenLabs, which are especially realistic compared to the models of the past you’re probably used to.3 This makes it really easy to listen to the report as I’m walking, on the bus, eating, doing low-intensity cardio, cleaning, or cooking. Even if I’m not occupied with a physical task, reading along with the voice helps me pay attention when I’m tired.
One major problem with this workflow is that it doesn’t involve any reflection or active recall, which is essential for integrating the knowledge into my long-term memory. Really I should be generating flashcards to import into my spaced repetition app (also RemNote) and/or set reminders to recall everything I can about each report a few days after I read it.
An alternative LLM-augmented voice transcription app is Superwhisper, but it’s not as polished yet.
To be fair, I haven’t had access to one!
Readwise Reader uses slightly outdated ElevenLabs models; for the latest ones, use ElevenReader, ElevenLab’s in-house text-to-speech app. But it’s buggier and doesn’t have all the features of a proper read-later app.