My current writing workflow.

As of Feb 2025

I’ve written about how I use language models to help with writing a few times already, but things are moving quickly and my process has changed enough that it seemed worth producing an update. These days, instead of typing first drafts, I usually start from a transcript of me talking out loud. I’ve found this really helpful for getting from "rough ideas in my head" to "something other people can read," especially for anything longer than a slack message. I originally started doing this kind of thing with ChatGPT’s voice mode, and while this did let me come up with the best blog title I’ve ever written, I’ve gone off this compared to just monologuing. The questions it asks aren’t worth the friction.

Why I find speaking easier

I think that this process helps because, especially if I’ve been thinking about something for a while, I don’t just have a single line of thought about it. It feels more like there’s a network of connected and overlapping threads all fighting for attention at once, and it’s hard to get them all out. I’ll start in one place, but realise that that place builds on something I haven’t said, but starting with that thing also doesn’t make sense…

When I'm talking, especially knowing there’ll be a transcript, I don't feel like each sentence needs to come out in a perfect order — I can be messy, backtrack, figure things out as I go, and decide to come back to things later. I can dump all of the information out of my head, and then worry about re-arranging it.

Despite the mess I’m describing of the ‘whole picture’, at a sentence-by-sentence level, I feel like I have the opposite problem. I want to very precisely convey the thing I believe, with the confidence I actually have. When I write, this means I often get stuck editing before there’s anything to edit — crafting the perfect sentence, deleting it, and starting over. When I’m speaking, I rely a lot on tone of voice to signal uncertainty or emphasis, and I still do this even when it’s for a transcript.

These markers get lost in transcription, but that's actually fine — getting the ideas out is what matters at this stage, and the emphasis I add lets me feel like what I’m expressing is true. If, having lost that emphasis, the words don’t convey the same feeling, I can add the nuance back in when editing. I don’t think it’ll be too long before models can pick up on my tone well enough to help with this stage as well, but for now I do it manually, and knowing that I can is enough to stop me agonising over the wording as I speak.

Setup

I record myself thinking out loud about whatever I want to write about. Sometimes this is a proper monologue, sometimes I’m responding to specific questions or prompts I've prepared (or asked Claude for). Occasionally I'll instead record a conversation with someone else and use that as the starting point, though this obviously requires someone else to be both available and interested in discussing it.

My current toolkit is pretty simple:

1. Fireflies for transcription

2. SuperWhisper for dictation

3. Claude for the bulk of my writing and drafting, other models for second opinions

4. Examples of my writing style (added as context to Claude projects)

This last part is important - I've found that giving more examples makes a huge difference in getting output that matches my voice. See this post for more.

Once there’s an initial draft or outline, I use SuperWhisper to give verbal feedback on drafts. I'd never ask a colleague to do multiple full rewrites based on quick verbal feedback, but when writing is instant and free, why not? I'll do several rounds of draft-and-feedback until it feels 'close enough', then move to Google Docs for manual editing (though I still use SuperWhisper for longer sentences).

I've gotten better at judging when a draft is ready for that final editing phase. Even with all my examples and feedback techniques, Claude's output still always needs tweaking - but that's fine. Getting 75% of the way there in half the time is a huge win.

Once I’m editing in a doc, if something doesn't sound quite right and I'm trying to work out how to rephrase it, I'll often go back to a model for suggestions. I tend to paste the whole draft I'm working on back into the chat window, and then just quote the section I'm trying to rephrase. I think this helps match the style I'm going for in the final piece. When I’m happy with the final output, I’ll get at least one model to proofread, though I might not listen to the suggestions.

Leaning into AI’s strengths

I think the way to get the most out of language models is leaning into the fact that they have a completely different cost profile from human assistants. Tasks that would be tedious or annoying for humans - like generating ten different possible structures for the same content, or reading through pages of messy thoughts to spot patterns - are trivially cheap for AI.

I think you can push this advantage really far, especially as it’s often much easier to pick from a list of options than describe exactly what you want. Here are some examples of things you can ask for:

• Multiple possible structures for the overall piece before moving from transcript to draft

• Different ‘opening hooks’

• Both of these combine e.g. I often ask for bullet-point outlines, each with a complete first paragraph to get a sense of the overall vibe

• Different ways to break up sections that feel too long

• Multiple rewrites, in different styles, to fix a section that doesn't quite work

• A list of clarifying questions (where I can then just answer the interesting ones, out loud)

These are all things I’ve actually done, but if you thought I generated all of them without asking Claude to suggest things I might have missed, you haven’t been paying attention.

Similarly, if I want feedback, I’ll often consult multiple models, or multiple ‘styles’ from the same model with different contexts, and have each generate concrete suggestions. If you’ve already written a feedback request prompt for one, the cost of pasting it into a different chat window is basically nothing. You don’t have to listen to every piece of feedback, treat the lists as “things to consider”.

Questions I’m still asking

I don’t know how big a deal the lost nuance from speech is, though I don't expect it to be too long before models can process natural language in a way that includes and accounts for inflection. I'm also still doing a lot of experimenting with which kinds of writing and which models end up being most useful. So far my default is Claude, though I've heard good things about R1 as a writer, as long as the writing isn’t sensitive!

My main plan for finding other improvements is to keep asking: what are the things models are unusually good at or find especially easy? How can I push further in those directions?

I expect my process is going to continue to change, and it already looks massively different to what I first described on this blog (less than a year ago). Maybe by 2026 I’ll be able to have a natural-feeling conversation with a skilled interviewer, and immediately turn it into the post I wish I’d written. But while there’s still any room for my own input, I expect that trying to understand and lean into models’ strengths will be the best way to amplify it.