I want to take a short break from our regular topics this week to talk about AI. In particular, I want to be up front about how I do and don't use AI to produce articles.
I write the first draft of these articles completely by hand with no AI involvement. I typically outline them first, using my notes in my main Obsidian Vault for reference and inspiration, but again without any AI input. Early revisions to the first draft are again done by hand, rereading and editing as I feel inspired.
I do use AI tools for editing and for technical tasks for publishing, as described below.
Editing
I write these articles in plain text Markdown files. Once initial revisions are done, I do typically use the Gemini CLI to perform an editing pass on them. The AI is instructed to check for:
- typos and misspellings
- grammar errors
- incomplete thoughts or paragraphs
- stylistic inconsistencies
While I have a style guide which provides guidance in all these areas, I have found it most effective to have it focus on one of the above at a time.
My experience is that it is very good and useful for the first three items above, it is always eager to tell me that my writing style is great and that there are no problems.
Publishing
This site is published by a series of custom tools which transform the raw Markdown source into proper HTML & a structured site. I used AI heavily to develop those tools - it's all stuff that I could have coded by hand if I wanted, but if you can accurately describe what you want it is so much quicker to "vibe code"1 them. In this case the scope is small enough and the stakes are low, so I'm not worried about scalability, long term maintenance across a large team, or similar concerns. At work I am much more cautious about dealing with LLM generated code, but here as long as I scan it to make sure it's not going to delete random files on my hard drive the potential downsides are pretty limited.
Similarly, I'll sometimes use AI to update HTML or CSS used in the final site. These are things that are further outside of my personal expertise (especially trying to make things look good) but are very low risk if it does something weird.
Summary
I've found these tools to be valuable for assisting around the edges. They can catch things that I am likely to miss and speed up technical tasks that realistically I would be unlikely to carve out time for. I deliberately don't use them in a way that "writes for me" or is even close to that. An idea that I've come across in several sources over the past year or so is that "Writing is Thinking". The short version is that you don't "think about stuff and then write it down", the process of writing is how you actually think. The act of working through getting your thoughts down onto (virtual) paper is what forces you to make them coherent. I've tried to embrace this, it has resonated and helped me. When I free myself up to just start writing something even if I don't know where it will end up, I can find myself getting into a flow state and letting the ideas and words come together as I go. My old pattern, where I try to think of everything and have it 100% in my head before I write anything down, is a recipe for endless procrastination because for anything moderately complex, you'll never have it all in your head ready to go.
As of the publish date, this accurately describes all the articles I have published so far. If future articles ever deviate from this process, I pledge to disclose the changes and, if needed, update this description.
Previous: Obsidian and Brindlewood Bay
Next: Docs and Lancer
Footnotes
-
Ugh in general. I'll link to Simon Willison for details, because I'm not gonna link to Twitter: https://simonwillison.net/2025/Mar/19/vibe-coding/ ↩