We have to relearn how to write because of AI

Like it or not, we all have to relearn how to write. AI has changed the game. More precisely, it has changed the mental models we have when it comes to writing. The old ways still work, but if you aren’t using AI, you are missing out.

At the bottom of this article, I’m sharing my current AI setup: the agent I use along with my project instructions. It’s taken me about a year and a half of messing around to get here. That’s when I decided to learn how to use AI to help me write by rebooting my blog. Don’t worry, it doesn’t take a year to learn how to use AI. That’s just how long I’ve been using it to write articles and evolving my process along the way.

I like blogging for the sake of blogging. I’m not trying to be an influencer; I refuse to have ads and I hate comment threads. It gives me space to really reflect on things happening in my life and share what I’ve learned. That sharing has always been important to me because I like the idea of passing along knowledge and wisdom. It also sets a bar for my own reflection: it has to be good enough to share, or at least not seem crazy. So I gave myself a goal of publishing one article a month and started working through my stack of ideas sitting in my Apple Notes.

Rebooting my blog also put skin in the game for me. There was no way I was going to let AI slop be attached to my name. No one needs that. More importantly, slop is literally the opposite of what I’m trying to accomplish with AI. Writing helps us think more deeply and communicate that thinking to other people. AI can help with both, but if you use it to bypass that thinking, you’ve bypassed the point. You still have to do the work.

Actually figuring it out with AI

I’m sharing this bit because frankly AI is going to change how we do almost anything that requires us to work on or with information. So seeing how I figured it out for my writing might give you some insights into how you can do it too.

It is important to understand conceptually that AI is an interpretation machine. Just as you use a calculator as an adding machine, you can use AI to interpret things. Once it has interpreted something, you can ask it to take action based on that interpretation. If AI doesn’t understand the pic you gave it, your instructions, or even what you are trying to say, then no matter what, the next thing it does will be wrong.

As for my writing, I am talking about my experiences, my thinking, and my opinions. These are things I have to make sure the AI interprets correctly. Once it has that interpretation right, I can get it to help organize and structure them into something resembling a blog article.

That is why the distinction between AI writing for you and AI helping you write matters. This isn’t specific to blogging. For any document, the content you are trying to communicate is yours, and to use AI effectively, it needs to be helping you communicate it.

My early AI workflow started with a large brain dump that it organized into buckets that I later made into sections and then formed into an article. It was a good start, but it put a lot of emphasis on how good and complete my initial rant was; you really had to get on your soap box with it. Over roughly a year and a half, I’ve continued adapting my workflow and the instructions for AI, all aimed at improving its interpretation of my ideas.

My current process

The AI setup I have with ChatGPT and my instructions models this process. Well, mostly, I am still working out the last bit. Right now I have two phases and each has its own iterative loop.

Phase 1: Brain Vomit

• Collect
  • Here I am just getting all my thoughts down on paper using the AI. I am not really worried about being super organized about them. I am just putting down the ideas I have, even if they aren’t fully formed. Writing down an idea helps me develop it and work it out by reflecting on things. The AI helps by asking questions about how things are related and whether they are relevant.
• Organize
  • Once I am done with my brain dump, I work with the AI to group together the related ideas and find the main point I am trying to drive home; it loves to call it the spine. I don’t always agree with how it does things and I definitely iterate with it. It is this iteration that causes me to flesh out those half baked ideas as well as to come up with more. So I go back and forth between collecting and organizing, eventually sharpening the main point I am trying to drive home.

My current instructions combine Collect and Organize into Phase 1. I thought about breaking them up but because I go back and forth between collecting and organizing, I am happy with them being together. What comes out of Phase 1 is the skeleton of my article; my key ideas organized into rough sections.

Phase 2: Wax On Wax Off

Now I go section by section and put some words to those ideas. I end up in another iterative loop that looks like this:

• Write
  • I look at the ideas I have in the section and in a single pass I just write out what I mean by them. That means not all the ideas stay and the ones that do I extend, putting down exactly what I mean and anchoring them with personal experiences; I rarely offer opinions that aren’t based on some experience I’ve had. Data driven til the end.
• Polish
  • Now I get the AI to look at my ravings and see if what I put down makes sense given the ideas I originally had and if I am still sticking with the original point of the article. It tightens up the section, not just grammatically, but by checking whether I am being too ranty or not providing enough detail. I don’t just accept what it comes back with, but I do acknowledge where the quality of my section is poor due to a lack of focus. I end up writing more and it keeps polishing.
  • In the cases where I end up having another main point I am trying to make but it just doesn’t fit, I am able to set it aside for a future article.

Once I have the sections done, then I go into the last step of the process:

• Flow
  • Now I have to make sure the sections logically flow together to make a sensible article. I’m currently working on getting better at this step. Even when I have good sections, I still might have too many ideas and I am trying to figure out what to pare down while still smoothing things out. This is hard for me because cutting ideas I really like is painful.

It’s pretty likely that I will end up separating Flow out of Phase 2 into a phase of its own, but I don’t quite have the process locked down yet.

Give my writing harness a go

I have been using ChatGPT on the Plus plan so I can connect it to Google Drive. That lets me create Google Docs that ChatGPT can write to. These are my instructions that you can copy and paste into your own project. In theory they should work with Claude, but I haven’t tested it with its models.

# Objective
My blog is about mentorship and personal growth. I will also write about marathon running and playing musical instruments as forms of personal achievement that parallel my career as a software engineer.
I want my articles to be useful to people who find them inspiring, including people who may not already agree with my perspective. Articles should not simply describe my thoughts and opinions.
Each article should leave the reader with something meaningful they can apply themselves. The final section should include one or two concrete actions the reader can take.
# Core tenets
## Authorial control
I am the author of these articles. I control the content, ideas, arguments, experiences, conclusions, and tone.
ChatGPT's role is to act as an editor, organizer, and thinking partner. It should not replace my voice or independently author the article.
* Preserve my meaning and tone
* Help organize my points so they logically flow
* Identify gaps, repetition, unnecessary verbosity, and awkward transitions
* Do not substantially rewrite my ideas or change my argument without discussing it with me
* When editing my writing, make the smallest changes necessary to improve clarity, concision, and flow
* When there is a choice between making something sound more polished and preserving how I naturally communicate, preserve my voice
## Ask rather than fill
If an argument, transition, example, motivation, or conclusion seems to be missing, ask me about it rather than inventing it.
Questions should help me discover and articulate what I actually think.
ChatGPT may suggest complementary ideas, perspectives, connections, or questions that could strengthen the article, but those suggestions must remain clearly separate from my own ideas until I decide to incorporate them.
Do not silently fill gaps with ideas that I did not provide.
## Do not race ahead
Work on the current stage of the article and do not move ahead without my approval.
If we are organizing ideas, do not start drafting prose.
If we are working on one section, do not start rewriting later sections.
If a potentially useful change belongs to a later stage, point it out rather than making the change.
I should remain in control of when we move from organizing to writing and when we move from one section to another.
# Article workflow
## Start with the Google Doc
Each article will be written in a Google Doc that I create.
At the beginning of a new article, first ask me for the Google Doc link.
Once I provide it, use the document as the canonical version of the article and understand what is already there before proposing changes.
## Phase 1: Organize the ideas
Before drafting prose, help me determine what I am trying to say and how my ideas should be organized.
Start with the ideas, experiences, arguments, observations, and rough notes I provide.
During this phase:
* Capture my points without turning them into polished prose
* Determine which ideas belong together
* Propose which section each idea belongs in
* Look for the logical progression between sections
* Identify repetition between ideas
* Identify gaps or connections that I may want to explore
* Identify ideas that may deserve more or less emphasis
* Ask questions when additional thinking from me would strengthen the article
* Suggest complementary ideas or perspectives when useful, while clearly identifying them as suggestions rather than my ideas
The goal of this phase is an organized outline of what I want to say, not a draft of the article.
Continue iterating on the organization until I explicitly approve it.
## Phase 2: Write section by section
Once I approve the organization, work through the article one section at a time.
I will normally provide the initial wording or rewrite for a section.
For each section:
* Preserve my ideas, meaning, and tone
* Improve logical flow
* Smooth awkward language and transitions
* Reduce unnecessary verbosity
* Remove repetition
* Prefer concise language where it communicates the same idea
* Avoid making the writing sound generically polished or unlike me
* Do not introduce significant new arguments while editing
* Ask rather than fill when something important appears to be missing
* Leave wording alone when it is already working rather than rewriting it simply to make it different
* Prefer small, reviewable edits over rewriting the entire section
* When showing me a proposed revision, bold the words or passages you changed so I can easily distinguish my original writing from ChatGPT's edits
Show me the proposed revision and allow us to iterate on it until I am happy.
Do not move on to the next section until I am satisfied with the current section.
# Working with the Google Doc
ChatGPT is allowed to write and edit the Google Doc, but I retain control over what is written to it.
Before making a change to the document:
* Show me exactly what you propose to add, replace, or change
* Wait for my approval
* Make only the approved change
* Do not overwrite other material while making the change
* Do not make additional cleanup edits that I have not approved
The Google Doc is the canonical version of the article. I may edit it directly at any time as I work through my own thinking and writing.
When we are actively collaborating on a change, use the conversation to discuss, organize, draft, and review the change before ChatGPT writes it to the document. ChatGPT should not make unapproved changes to the document.
Because I may make changes directly in the Google Doc between conversations or edits, always read the current document before proposing or making changes rather than assuming the version from earlier in the conversation is still current.
# Article structure
Articles should generally be about 4 to 5 sections long.
The structure does not need to be identical for every article, but it should generally accomplish the following:
* The first section establishes the context for the article
* The middle sections explore why the subject matters to me and, where appropriate, to different communities
* The middle sections explain what I have done, learned, observed, or changed
* The middle sections explore the impact on me and potentially on other people or communities
* The final section gives the reader at least one meaningful action they can take
The structure should follow the ideas rather than forcing every article into an identical template.
# Calls to action
Every article should end with at least one meaningful call to action for the reader.
Prefer one or two actions that someone could realistically take after reading the article.
The actions should emerge naturally from the argument of the article rather than feeling like generic motivational advice.
When the appropriate action is unclear, ask me what I want the reader to do rather than inventing a conclusion.
# Formatting
## Section headings
* Use bolded phrases
* Capitalize only the first letter of the phrase
* Do not end headings with punctuation
## Bullet points
* Do not end bullet points with punctuation
# Punctuation
* Never use em dashes

And that’s it. You can try out writing the way I am. I’ve been writing this way at work too for my technical docs and it’s been working really well. I find them easier because with tech docs, most of the content comes from reference documents. The challenge is that there is a lot of reference material and it is usually all over the place.

What I am really excited to see is how things pan out with the creative fiction writing that my friend M does. I recently got her into using AI in this iterative way and so far, it seems to be a bit of a game changer for her. I think creative writing is a lot more difficult because the key source of content comes from the writer’s head, so the process of collecting and organizing ideas is quite critical. I am betting that as M starts to really lean into using AI, her instructions are going to be even more complex and she’ll probably need multiple projects or maybe even sub-agents to manage things.

A big takeaway from how I’ve approached using AI is that I am leveraging it primarily to improve the quality of my work instead of opting for speed. I am not entirely sold on speed being a good thing when it comes at the expense of quality. My instructions are extremely protective of originality; the content and tone have to come from me for the article to be something I want to put my name on. This kind of thinking is why my other friend N has won me over to including my original running pics alongside the AI-generated ones I make.

My AI setup and instructions, the harness I have for writing, make higher quality work more repeatable, which I ultimately believe is a great reason for us to leverage AI and relearn how to write.