Writing with AI, Responsibly
Starting with this issue, I’m using AI to help me write this newsletter. Please don’t stop reading if this is upsetting–I think I can win you back. Allow me to explain how I’m using AI in concrete terms, and then address head-on what I imagine will be your objections.
First and foremost: None of the words in this newsletter will be written by an LLM. They will remain my words and my words only.
With that clarification, what has changed is that I’m now using AI as a writing coach.
The skill I’m using does this by keeping its own separate markdown file in which it captures everything it understands about what I’m trying to say. It then challenges any ambiguity in my writing, any missing arguments or lack of rigor, and pushes me to correct these issues in my own words.
The impression I get when hearing people talk about using LLMs for writing is that they believe it to be inauthentic, lazy, and error-prone. The workflow I’ve described is intended to achieve the opposite on all counts.
Authenticity: I’m committed to preserving the connection between my mind and yours, mediated only by my words. For this reason, I won’t allow my AI workflow to contribute its own words, but only to challenge my writing and thinking, leaving the correction of the issues it identifies to me.
Effort: This workflow makes writing this newsletter take significantly more time than it has in the past (confirmed, as I’m using it to write this issue!). It forces me to engage more deeply with what I think and what I’m trying to say. That engagement takes more time, more effort.
Accuracy: The AI’s role is to challenge where my writing is poorly thought-through or communicated. It does not provide its own information or evidence, side-stepping the issue of AI hallucination.
To this point, what I’ve published in this newsletter have been stream-of-consciousness first drafts. This has allowed me to tiptoe back into writing when doing anything else would have been anxiety-inducing for me.
But I don’t want to leave my writing at this point. I want to begin pushing the quality of what I publish, and I hope that this new workflow will allow me to achieve this.
Here’s the skill I’m using if you’d like to see exactly how it works and perhaps try it for yourself.
Instructions for using skills with various tools:
I tried the skill (via Claude Code) on one of my own blog posts, and it's very blunt about what it doesn't like! You don't have to worry about it being overly sycophantic, that's for sure.
One of the reasons why I've been hesitant to use these AI writing aids is that I feel like taking its suggestions will make my writing more generic and 'LLM-ish', even if it might get to the point quicker. Maybe this is a better way as it makes you write out its suggestions in your own words rather than writing it for you?
Amen to this. Just the other day I found https://www.roastmypost.org/ which looks promising. I like Christopher Moravec's way of putting it: Don't be a secret cyborg.
Here's my own take on AI-assisted writing from a more innocent time (this past November): https://agifriday.substack.com/p/writing