Ai-enhanced writing process

In the last six months, AI tools have changed not just how I code, but how I write. Recently Gwern pointed out that I was letting AI write for me, defeating the purpose of writing in the first place. I reconsidered what these tools are actually useful for and how they can enhance my writing instead of getting in the way.

So I came up with two principles:

I collect ideas

For a long time I’ve used Notion to track my ideas and organize the process of writing, editing, and publishing. None of that has changed with AI. I continue to come up with new ideas that I add to notion. When a particular topic seems to have reached a critical mass of detail, I bring it into my writing queue.

AI rearranges the outline

This is where AI has started to help quite a lot. Without AI, I used to sit for a long time moving thoughts around, trying to figure out how everything connects.

AI tools make this much easier now. I can copy-paste the raw notes into Claude and ask for variations: “Can you structure this into a three-part story?” or “Can you arrange this chronologically?” I keep prompting and tweaking until I have an outline I like. Then I ask for a final version, making sure it doesn’t lose any of my original ideas or add any of its own. The outline is still entirely my thinking — all the LLM has done is moved things around.

I set up the draft

With the outline in hand, I open a fresh conversation with Claude. I paste in the finished outline, ask it to create an artifact, and give it a simple set of instructions: I’ll be pasting in speech-to-text transcripts section by section, and I want it to clean each one up — fixing incorrect words, spelling, and grammar — without making any significant changes to the text. Because Claude has the outline I wrote by hand, it does a good job of correcting words that got misheard. This comes up with proper nouns which speech-to-text often mangles.

I dictate, AI cleans up

Then I open Spokenly, a speech-to-text tool that allows me to choose between local and cloud-based models for transcribing speech. I use the ElevenLabs API, which offers a generous free tier and appears to be the most accurate.

With Spokenly open, I look at the outline and speak in a natural voice, as if talking to a friend. This is the fastest way for me to turn thoughts to words and allows the blog to feel conversational. Once a section is done, I wait for Spokenly to output the text, paste it into Claude, and rinse and repeat until I have a complete draft.

AI critiques the draft

With a draft in hand, the next step is to critically evaluate whether the words I’ve spoken actually convey what I was trying to say at the outline stage. I formerly used a chat interface for this, but I found that hearing the ideas spoken back to me is far more productive — it keeps me from getting stuck on small details like word choice, grammar, or the way things look on the page.

Here I use Google NotebookLM. One of its features is a podcast where two hosts discuss the source material. I specifically use a critique podcast, which I’ve only found in the desktop web version. The two hosts jump straight into critiquing the ideas and suggesting improvements

About half the criticism they raise are things I wouldn’t have thought of immediately — ideas a human editor might surface — and they all push on how effectively I’m conveying my ideas and building my arguments. The other half tend to be a bit wacky, creative but not exactly useful.

That mix is what makes this process productive. It feels a lot like getting a human edit, where the feedback isn’t always perfectly aligned with what I want. It doesn’t have the sycophantic quality of a chatbot that tells you everything is great.

I edit by hand

From there, I write a short list in Notion of the structural changes I want to make, and then I go through and do all of it by hand. Writing by hand forces me consider the high-level ideas and how they translate at the sentence level. It also keeps me practice critical thinking and writing, which are ultimately why I write in the first place.

Finally, I create a new branch, bring the post into my codebase, add visuals and any supporting materials, and open a pull request. Netlify creates a preview build that I send to my wife for the first human edit.

An evolving process

This process will very likely change given how quickly the tools are evolving. So far I have focused on what I do best — thinking, connecting ideas, writing the actual words — and letting the LLM do what it currently does well — transforming information from one format to another, rearranging ideas, offering critique, and yes, sometimes hallucinating in ways that turn out to be useful.

The tools feel like amplifiers now. And this process leaves me confident that I’m not falling into the trap I fell into before, where I was letting AI do the thinking and writing for me.

---

Thanks to Q for reading drafts of this.