Recently, I’ve been thinking deeply about different ways to use AI to help people become more effective and productive. I’m doing this for work. I recently joined a startup called Enhance Labs and our goal is to develop a tool that makes you AI native (or something like that, it’s a startup, it changes every day).

Dictation

A pattern I’ve been exploring, and am currently right now to right this blog article, is dictation. The idea is that speaking allows thoughts to flow quickly. You can capture a true stream of consciousness much faster than writing.

Dictation isn’t new, obviously. 10 years ago, at a law firm, I saw lawyers dictate their thoughts into microphones. Paralegals would listen to these recordings and convert them into something, I don’t remember.

Now it’s different. LLMs have the ability to immediately process these thoughts without another person’s assistance. After dictating, I can get AI to clean up the content and critique the overall narrative. Then I sit here like I am now, editing it, thinking about it and getting the same benefit I usually get from writing (that is, refactoring the thoughts in my head).

I’m not entirely sure it’s a positive productivity gain. I’m really scared of LLMs creating a society with too much content. But for the moment, I’m finding it useful. It’s making me think more.

Speak more, be better spoken

Writing is good for clarity of thoughts. I think speaking helps there too, and of course, makes you a better speaker. I think speaking probably helps you improve TTFAT (time to first articulable thought). I’m too early though to note anything here.

The workflow

1. Record dictation.
2. Proofread manually to check for words not transcribed properly.
3. Dump into an LLM to clean the transcript up (but make sure it doesn’t change what you said, that’s lame).
4. Read it, edit it, think about it, decide if it’s a thought you like

I’ve been experimenting with this workflow for a number of prompts (brainstorming, extracting actually useful notes from meetings, etc.)

My prompt

This is the prompt I use for transcript cleanup. This plays with the ordering of the content as well, and tried to condense it down. I’d recommend keeping the condensed bit in. It’s very interesting to see when you talk to your computer for 15 minutes and it summarises your thoughts in 2 sentences.

You are a professional transcript editor.
Your job is to polish my raw dictation so that it’s clear and easy to read, without changing the meaning or adding any words I didn’t say.

My raw transcript. Clean it up by:
1. Removing filler words (um, ah, you know, etc.).
2. Correcting grammar, punctuation, and spelling.
3. Keeping my original phrasing—do not introduce ideas or sentences that aren’t in the text.

Output 3 sections:
- Cleaned transcript: Output **only** the cleaned transcript (no commentary).
- Narrative-adjusted: Output only the content/ideas from the cleaned transcript, but reorder the content in order to create a coherent narrative.
- Concise: Take the narrative-adjusted version and condense it down as much as possible into just the core ideas.
	- It should still be full sentences, but every sentence should be dense with information.
	- You might turn 1000 words into 20, there is no constraint you need to follow on word count.
	- Only thing that matters is that every word written is not wasted

Output should have the following format:

---
## Cleaned Transcript

<content here>

## Narrative-Adjusted Transcript

<content here>

## Condensed Transcript

<content here>
---

Here is the transcript:

## Transcript:

I think speaking naturally to computers will become more common now. It feels a bit weird still, but human behaviour will change over time. Dictation has me engaging more with my thoughts, so it’s probably a step in the right direction towards whatever ‘AI native’ evolves to be.