My Writing Workflow

Sharing how I am writing posts these days – being open about how I'm using AI to help edit and structure my words but not generate them.

One of the things that has changed over time is how I approach writing for the blog. I've tried a variety of approaches, partly because writing doesn't come naturally to me – but I value what it does. Getting the ideas out of my head and somewhere else creates space and a sense of mental clarity that's genuinely good for me. That clear space allows me to move on and think about other things rather than replaying the same old loop.

At the very beginning of my blogging life (yes, I still need to migrate my archives across to here), most of what I wrote was essentially automatic writing. Stream of consciousness, write & hit post. The words were sometimes problematic: grammatical errors, spelling mistakes that slipped through. That editorial component of writing is something I've always struggled with. Over time, I've adapted tools to address those shortfalls because I know where my gaps are. While it would be great if this were a collaborative process, I actually enjoy being edited and having another person involved to get feedback on my writing. But I've always wanted blogging to be an independent process and not rely on anyone else, so I've used tools along the way to help fill the gap.

One early adaptation was to speak into my phone instead of typing; then I'd edit the transcript and, through that, develop a blog post. It was faster in some ways, but I didn't enjoy the editing – and that's what I ended up spending all my time on! I started to use Grammarly to help with the editing because it does a great job of all the things beyond spellcheck that I need, and I've stuck as a paid user since[1].

Then I moved to adopting an outlining process for posts, first developing a dot-point structure and coming back to fill in the details. That worked OK, and I still use it, but it lacked the organic feeling I would get from writing automatically or from speaking and transcribing.

I've come back to the transcription method, and a couple of pieces of software have really changed how I work. The is MacWhisper – a relatively inexpensive app that runs a local language model and produces remarkably good transcripts. I've tried most transcription tools as they've come out over the years, and each has had their pros and cons. MacWhisper isn't flawless, but it's pretty close and handles my accent well, manages my tendency to mumble through words. What it does a really good job of is resolving the ambiguity that comes with spoken English – where words sound similar, but only one makes sense in context. The result is still a transcript that needs editing, but the editing is much cleaner – less "what did I actually say here?" and more straightforward editorial work.

Most recently, I've added Claude to the mix – not for generating posts, but for the editorial pass. There is news this week from Anthropic about adding a watermark to text generated by the tool, which is technically quite interesting, but it's also caused quite a stir. I'm not sure if or when the tools to detect its use will arrive, but I thought I would be open about it and share my experience and how I am using the tool.

So my workflow is pretty much the same:

There are a few steps in there and a bunch of tools. But for me, they each serve a specific purpose and role. Throughout all of this – these remain my ideas. The words, well, they've changed and evolved throughout the entire process. What I feel makes them mine is hitting that publish button and taking responsibility for them being in the world.

For me, the role of the LLM, the AI, isn't present in the idea, and it's not in the words themselves. I assume that a watermark may say differently – but I do wonder if a pass through an LLM means that the content was in any way AI-generated. I think that distinction matters. There wasn't a prompt: write a blog post about this topic. The use of an LLM was a single step in a broader process: "Here is a blog post I've already spoken; now help me with the formatting and structure I haven't had time to do." For me, the LLM has a distinct value proposition, and it plays to what these tools are actually good at – processing text, applying pattern recognition and finding structural inconsistencies. And these are genuinely things that are not a strength of mine.

I can understand those out there who write, and many who write for a living, who find the use of LLMs abhorrent. I do find the generative use of LLMs quite problematic, and the output to be mostly just shit. But I don't feel my use falls into that same basket. The LLM is used as step in a feedback loop – it gives me something to work from, not start with. It's not a replacement for my own judgement, and I never intend to outsource that.

It's still me clicking the buttons and making decisions throughout the process. It's still my words, processed and refined yes, but still mine.

  1. Somewhat reluctantly! If anyone has suggestions for cheaper or better alternatives, please let me know!! ↩︎