How I taught ChatGPT to help me write, without letting it write anything
When I say “article mode,” you say “productivity”
I’m dictating this article. When inspiration strikes, I want to start writing right away. I never thought I’d be a person who dictates articles, but I’m currently on a walk, and I wanted to share thoughts on how I use dictation to compose articles in ways that I find surprisingly powerful and robust. (And now that I have a site that I feel a bit obligated to feed with articles on a regular basis, I’m extra grateful that I found this tool!)
But first, a quick aside:
I had a friend in college who was a couple years behind me. Because I’m old, I went to college in the era where everybody traded their AOL Instant Messenger screen names. My new friend had created his at college, and his username was WhenISayPunchIs.
I couldn’t understand the meaning of that username. Eventually, I asked him, and he reminded me about a story I remembered vividly from my own freshman orientation, when the student dean — a charmingly hilarious woman named Alwina — gave a talk to the entire incoming class, warning them about important topics, including: dangerous alcohol intake.
That dean started a chant across the entire first-year class.
“When I say punch is, you say evil! When I say punch is, you say evil!”
Her point was that you don’t know exactly what’s in punch, so you should probably not drink it.
A reasonable lesson, communicated in a silly but clearly memorable way. So my pal Jonathan took WhenISayPunchIs as his AIM username. Okay.
This is a very bizarre way of getting to my point. But if you’re like me, you might on occasion have the nagging sensation that some people are using various technical tools in ways that are more savvy or clever or powerful than you are, and that could benefit you to learn about.
I’ve written before about how I use Claude to proofread every post that goes live on this site. It’s automated, and I don’t even have to think about it.
But I wanted to share a way that I use ChatGPT to help me when motivation strikes to compose an article and I am nowhere near a real keyboard.
I’m unwilling to tap out article-length content on my iPhone’s screen. But I am willing to dictate. I don’t find that Apple’s Siri-powered transcription is powerful enough for my needs. But ChatGPT’s transcription is incredible.
So I launch the ChatGPT app on my phone, and I tap the microphone button.
And before I started talking, I kept giving ChatGPT the same semi-coherent prompt each time I started dictating a story. I would tell the service:
I am going to be dictating to you and then submitting a big wall of text to you. Correct any obvious mistranscriptions, fix any grammar, and paste me back raw Markdown of the article that I have dictated. Don’t otherwise change my wording. Do honor any notes that I’ve given you inline.
But I gave ChatGPT that series of instructions every time I wanted to dictate an article.
So one day, recently, as I started a piece of dictation, I told ChatGPT that I was going to refer to this as “article mode.” And from now on, anytime I said “article mode,” ChatGPT should know that I wanted it to follow those instructions.
Because these services build themselves memories — really just text files full of notes about how you want them to interact with you — it was able to internalize that instruction.
So now when inspiration strikes, I tap the microphone button in ChatGPT, declare “article mode,” and just start talking. That’s how I started this very piece, which I’m still dictating live now.
I’m including a screenshot here so you can see the raw text I dictated to ChatGPT — including its mis-transcriptions, a moment where I asked it to link out to my story about using Claude for copyediting, etc. The result it spits back doesn’t change my words or rewrite me; it fixes my mistakes and adds formatting. It’s remarkable.
I wonder if this brings my opening story into context for you a bit. When I say punch is, you say evil. When I say article mode, ChatGPT knows I want a cleaned-up, formatted transcription.
What’s extra powerful is that because I do empower ChatGPT to follow instructions I toss out mid-article, I can have it find links or handle rephrasings that I come up with while talking.
I understand that some folks get irritated by using these AI-based tools. And as I’ve explained on this site before, while I won’t use them for farmed-out creativity, I do appreciate their ability to drive improved productivity, which is precisely how this usage works.
When I say article mode, I don’t think that’s evil.