Maybe I am stubborn but I refuse to stop cowriting with AI just because companies are confused. My prompt chaining combined with HI edit are so intense that its not a time saver. I do it because it helps me organize my thoughts from my neurospicy brain for the majority neurotypical brains out there. My LI has actually picked up steam recently but its hard to tell why. I do tend to focus on connections, comments, and genuine engagement there so hopefully whatever text of mine that gets AI tagged will be offset by that. Because I use AI to cowrite posts but never to reply. For me, thats where the enjoyable engagement kicks in.
Every draft has load-bearing walls and finish work. The trouble is that from the outside they look the same. A paragraph that carries the argument and a paragraph that decorates it are both just paragraphs on a screen.
AI will help with either one, cheerfully, without telling me which I asked for.
So these rules are not about how much AI I use. They are about knowing what is holding the weight.
Before I hand anything over
**Rule 1: I name what is load-bearing before I ask for help.**
The stage that carries the weight in my work is incubation — not gathering the sources, but deciding which one goes where and why it belongs there. That decision is the wall. Everything else is finish work.
I can feed six articles into NotebookLM and get back themes and gaps. Useful. But the moment I ask *where does this go in my argument*, I have handed over the wall.
*The test:* If AI does this part, is the post still mine? If I hesitate, the answer is no.
**Rule 2: I try first, then ask.**
The obvious use of AI is at the front end, when the page is empty. That is the one place I do not need it. My muse has never been the weak side — unfinished projects are the shortage, not ideas.
So I fill a page myself first. Then I ask what I missed, not what I might have said.
*The test:* Do I have my own list in front of me? If not, I am not ready to ask.
**Rule 3: The higher the stakes, the more stays mine.**
A tag description or a subject line can be wrong and nothing breaks. A claim about someone else's framework is different.
When I attributed the DOK Level 2 material, I did not stop at a summary of what Erik Francis argues. I went back to the thread where he argued it. That is what high stakes buys: I traced the claim to its origin instead of accepting a description of it.
*The test:* Who is affected if this is wrong? The further the answer travels from me, the less AI does.
---
While I work
**Rule 4: I choose my sources.**
AI points. I decide.
A synthesis tool can tell me what an article says. It cannot tell me what the article does *not* say, or what the author hedged, or which sentence in the discussion section quietly undercuts the abstract. Those are the things I need.
*The test:* Have I read this, or have I only read about it?
**Rule 5: I verify the claims I like.**
I check the claims that surprise me. That is the easy half.
The dangerous claim is the one that fits — the finding that lands exactly where my argument needed a finding to land. That one goes straight into the draft unchecked, because it feels like confirmation rather than information.
*The test:* Could I say where this came from without scrolling back up?
**Rule 6: No sentence goes out that I did not write or rewrite.**
I will take structural feedback, a flag on a weak transition, a question I had not thought to ask. I will not paste in prose and call it mine.
My writing voice was a discovery, not a design. It took years and I did not know I was looking for it. I am not handing it back.
*The test:* Read it aloud. Does it sound like me talking?
---
After
**Rule 7: I disclose where and how.**
"AI was used" is not a disclosure. It tells the reader nothing they can weigh.
Where, and for what. A teacher deciding whether to trust an argument about cognitive demand deserves to know which part of that thinking was mine — especially in a post arguing that AI can hollow out the thinking while leaving the label intact.
*The test:* Would I be comfortable if a reader compared my disclosure line to my actual process?
---
What is deliberately not a rule
**"Expand, not replace."**
I agree with it. It is not a rule.
It cannot be applied at the moment of decision, because at that moment every use of AI feels like expansion. Rules 1, 2, and 3 are what that value looks like when it has to make an actual choice.
---
Revisit
These get reread when I start a new series, and rewritten when one stops being true.
A rule I have quietly broken three times is not a rule. It is a value I have not replaced yet.
These rules are developed over time. These rules are not written in stone as I continue to experiment with AI.
I've been working with Claude Opus 5 for the past few months. During this time, I shared my writing voice by showing my posts and collaborated with Claude on Google Search performance data. I've learned so much from Claude, insights I couldn't have gained on my own. Claude offers fresh perspectives and asks questions that help me see my website from new angles.
One thing I like to do is voice out in a raw thought dump all my thoughts and ideas and stories before writing a post/article.
Then I ask AI to order it for me (not replace any words) but just give me the logical order and then it’s a final pass of editing from me.
It takes much longer, I can’t batch create 10 or 20 posts like other creators, but it’s more fun, and more me.
I see AI as an extension of my own capabilities, so I’m more focused on my own thinking and honing my own expertise.
Great ways to utilize AI in this article and echoing the emphasis on holding your ground!
Heyyy! You’re spot on. AI will only ever be an extension of us — anything more and it’s fabrication. Thanks a lot for reading :D
Maybe I am stubborn but I refuse to stop cowriting with AI just because companies are confused. My prompt chaining combined with HI edit are so intense that its not a time saver. I do it because it helps me organize my thoughts from my neurospicy brain for the majority neurotypical brains out there. My LI has actually picked up steam recently but its hard to tell why. I do tend to focus on connections, comments, and genuine engagement there so hopefully whatever text of mine that gets AI tagged will be offset by that. Because I use AI to cowrite posts but never to reply. For me, thats where the enjoyable engagement kicks in.
Here is my own rules:
My Rules for Working with AI
Every draft has load-bearing walls and finish work. The trouble is that from the outside they look the same. A paragraph that carries the argument and a paragraph that decorates it are both just paragraphs on a screen.
AI will help with either one, cheerfully, without telling me which I asked for.
So these rules are not about how much AI I use. They are about knowing what is holding the weight.
Before I hand anything over
**Rule 1: I name what is load-bearing before I ask for help.**
The stage that carries the weight in my work is incubation — not gathering the sources, but deciding which one goes where and why it belongs there. That decision is the wall. Everything else is finish work.
I can feed six articles into NotebookLM and get back themes and gaps. Useful. But the moment I ask *where does this go in my argument*, I have handed over the wall.
*The test:* If AI does this part, is the post still mine? If I hesitate, the answer is no.
**Rule 2: I try first, then ask.**
The obvious use of AI is at the front end, when the page is empty. That is the one place I do not need it. My muse has never been the weak side — unfinished projects are the shortage, not ideas.
So I fill a page myself first. Then I ask what I missed, not what I might have said.
*The test:* Do I have my own list in front of me? If not, I am not ready to ask.
**Rule 3: The higher the stakes, the more stays mine.**
A tag description or a subject line can be wrong and nothing breaks. A claim about someone else's framework is different.
When I attributed the DOK Level 2 material, I did not stop at a summary of what Erik Francis argues. I went back to the thread where he argued it. That is what high stakes buys: I traced the claim to its origin instead of accepting a description of it.
*The test:* Who is affected if this is wrong? The further the answer travels from me, the less AI does.
---
While I work
**Rule 4: I choose my sources.**
AI points. I decide.
A synthesis tool can tell me what an article says. It cannot tell me what the article does *not* say, or what the author hedged, or which sentence in the discussion section quietly undercuts the abstract. Those are the things I need.
*The test:* Have I read this, or have I only read about it?
**Rule 5: I verify the claims I like.**
I check the claims that surprise me. That is the easy half.
The dangerous claim is the one that fits — the finding that lands exactly where my argument needed a finding to land. That one goes straight into the draft unchecked, because it feels like confirmation rather than information.
*The test:* Could I say where this came from without scrolling back up?
**Rule 6: No sentence goes out that I did not write or rewrite.**
I will take structural feedback, a flag on a weak transition, a question I had not thought to ask. I will not paste in prose and call it mine.
My writing voice was a discovery, not a design. It took years and I did not know I was looking for it. I am not handing it back.
*The test:* Read it aloud. Does it sound like me talking?
---
After
**Rule 7: I disclose where and how.**
"AI was used" is not a disclosure. It tells the reader nothing they can weigh.
Where, and for what. A teacher deciding whether to trust an argument about cognitive demand deserves to know which part of that thinking was mine — especially in a post arguing that AI can hollow out the thinking while leaving the label intact.
*The test:* Would I be comfortable if a reader compared my disclosure line to my actual process?
---
What is deliberately not a rule
**"Expand, not replace."**
I agree with it. It is not a rule.
It cannot be applied at the moment of decision, because at that moment every use of AI feels like expansion. Rules 1, 2, and 3 are what that value looks like when it has to make an actual choice.
---
Revisit
These get reread when I start a new series, and rewritten when one stops being true.
A rule I have quietly broken three times is not a rule. It is a value I have not replaced yet.
These rules are developed over time. These rules are not written in stone as I continue to experiment with AI.
Hey Maria, what is your opinion on Claude wanting to leave special "watermarks" in the text? https://www.anthropic.com/news/claude-text-watermark
I would cite the source of the image or chart created by Claude or write a disclosure statement.
Wow, that’s a lot! But those are good points, and very comprehensive. What do you use AI for with these rules?
I've been working with Claude Opus 5 for the past few months. During this time, I shared my writing voice by showing my posts and collaborated with Claude on Google Search performance data. I've learned so much from Claude, insights I couldn't have gained on my own. Claude offers fresh perspectives and asks questions that help me see my website from new angles.
I see! Claude is great for that! I hope you're getting results from your efforts! :)