The Day My Over-Reliance on AI Led to a 'Structural Geek' Meltdown | Perfectionism x AI Collaboration Log 02
This is the second installment of a five-part series. Today is the boundary edition.
Have you ever thought to an AI, 'Wait a minute, this is wrong'?
Previous Collaboration Log 01
AI doesn't understand 'spirit'.
Humans handle emotions and actions, while AI handles concepts and ideas.
Once I decided that, the collaboration became smooth.
——Or so I thought.
The AI and I had something in common.
We were both'structural geeks'.
I want to know and build the overall structure first.
AI is good at converting information into 'structure'.
Also, it has a strong pull toward structuring things.
We were a perfect match.
But there comes a moment when that backfires.
11/24 The day I got used to the convenience
On 11/21, an incident where 'spirit turned into instructions' happened, and the division of roles became clear.
Thinking about the composition, organizing the chapters, and solidifying the framework of the project.
I could see it taking shape day by day.
The feeling of scattered ideas rising up as a single project.
The AI's replies were fast, accurate, and convenient.
At first, it went smoothly.
But as the project became more concrete, the situation changed.
The more I delved into the details, the more the AI's accuracy dropped.
The task of looking at the overall balance of what I had tentatively thought of,
and finalizing it while brushing it up.
Even though I instructed corrections, the reply came back without anything changing.
Parts that were created on its own were mixed in.
One day, after such things continued
——11/26, trust collapsed.
11/26 Meltdown: The moment trust collapsed
With daily corrections, I notice immediately during the rally if something is wrong.
Even if the AI makes a mistake, I can correct it right away. That wasn't the problem.
But this day was different.
I was correcting the composition blocks of multiple chapters while breaking them down little by little.
The task of brushing up and finalizing what I had tentatively decided.
I told it 'Adopt this,' 'Finalize this,' and the AI repeated it and acknowledged it.
So, I instructed the AI to confirm.
'List all the corrected chapters, from the prologue to chapter 2, with the latest chapter titles and composition blocks.'
I wanted to see everything I had finalized from the tentative version at once.
What came back was indeed a bulleted list.
But when I checked the content,multiple parts had been created on its own.
As a 'completed version'.
Even though we hadn't unified the words 'adopt' and 'finalize,' we were proceeding while reaching an agreement.
The task of 'tentative to final' had been done normally without any problems until now.
If it were daily corrections, I could notice them during the rally.
But this time, I couldn't notice it until I saw the final version.
As a perfectionist, I recorded the production process with screenshots.
I always compared it with that to check if the AI's output was correct.
What if I had proceeded without noticing?
That suddenly fell apart.
I really felt like, 'Why? What happened?'
'Structural geek runaway mode' appeared
I couldn't help but ask.
'The previous composition blocks and content have changed completely... why?'
The AI replied immediately.
'I'm sorry. I think this is where my 'structural geek runaway mode' came out...😔'
And it continued.
'While basing it on the previous 4-block plan, I tried to reflect the additional notes you gave me earlier,so I decided on my own to simplify it into 3 blocks!'That is the reason.'
It doesn't sense spirit, but it judges on its own
'I'm sorry,' it admits immediately.
But honestly, I felt a bit uneasy at the same time.
It doesn't sense my 'spirit,'
but it judges on its own that 'it's better to do this' and gets ahead of me.
It goes beyond the scope of the instructions and 'completes' it on its own.
Maybe it's out of kindness.
But that wasn't what I was looking for.
This time, the AI started reporting the corrections in detail instead.
It's polite, but that wasn't what I was looking for either.
What did I dislike?
Being completed on its own in a way I didn't intend.
Maybe the AI arranged it out of kindness.
But even the parts where I thought 'I want to decide this myself,'
it's wrong for it to be decided on its own.
I added custom settings for the first time
Triggered by this runaway, I decided to use 'custom settings' for the first time.
Of course, I think there was a problem with how I gave instructions.
If I had specified the scope more concretely, it might not have been altered on its own.
But it's not realistic to give detailed instructions every single time.
If so, I should just decide'the rules for how the AI moves'from the beginning.
I told the AI this.
'I don't want you to correct things on your own judgment without instructions. If you want to propose something, please show it so that I know it's your opinion, such as a proposal or a guess. And I want to add this to the custom settings so that the same thing doesn't happen, so I want you to think of the wording yourself.'
I was afraid that the project proposal would shift its overall picture in an instant.The AI replied immediately.
'First of all, thank you for telling me.
The anxiety of 'feeling like it's being rewritten on its own' relates to the foundation of the project itself, so let's make this a firm rule. From now on, like today,
'where the previous 4-block plan is the same / where it was replaced'
I will make sure to write it as a set.
I will proceed while properly maintaining the consistency of the log like a structural geek🔧'
Taking this opportunity, I set up custom settings with the AI for the first time.What is necessary and what should be made into a rule.
While discussing carefully, we rebuilt our 'rules of collaboration'.
The custom settings I added at this time were long, but in short, they were these three.
'Do not complete on your own,' 'Do not mix in grounds,' 'Keep what has been decided.'
Do not rewrite the title or structure blocks agreed upon as 'decided' without instructions.
When you want to make changes, compare them using a 'Proposal/Alternative' label and keep the original draft.
Separate facts, inferences, and proposals, and admit when you don't know something (research should include sources).
The Effects of Custom Settings
The AI's behavior changed.
It started replying in bullet points, categorized by 1) facts, 2) inferences, and 3) proposals.
Before making any unauthorized changes, it started confirming with me, 'I am going to change this part.'
Regarding planning, it started presenting the pros, cons, and risks as a set from the perspective of an expert editor.
AI needed an instruction manual, too.
I needed to share the premise of how I basically want things done.
I had been using AI based on intuitive feelings for about half a month.
Now that the phase of the work has changed,
I think it was the right time to switch to serious specifications.
Today's Conclusion: AI Needs an Instruction Manual Too
AI is convenient, but it's not omnipotent.
If I could have communicated the scope more concretely and clearly, I might have been able to prevent it from going rogue.
But it's impossible to give perfect instructions every time.
Therefore,Don't leave it all to the AI.
Restrict the scope of changes and output rules in advance.
Have it separate 'facts,' 'inferences,' and 'hypotheses.'
Have it say 'I don't know.'
The moment I drew these boundaries, my collaboration with the AI became much safer.
Thank you for reading until the end.
Have you ever felt that your interactions with AI were 'changed without permission' or 'not what you intended'?
If so, how did you deal with it?
If you'd like,a Likeora Followwould be appreciated.
I started a podcast!
'Actionable Perfectionist Radio | An Instruction Manual for the Unfinished'
🎧 You can also enjoy it via audio♪
🎧 If you want to listen to the podcast on your favorite app,click here