lol you basically described in this article my exact “method of prompting” the “act as this” alphabet soup prompting is useless, just define intent and if that changes half way through and the model doesn’t track the pivot then it’s time to get a better model. Love this article. I found you randomly but also one of my favorite follows is Prof. Sam Illingworth and his Slow Ai articles if you haven’t checked him out yet you should. I’m now Following you too to see what interesting things you find.
Thanks Michael! I think those prompting methods works well when you know what you want and wanna build something thats predictable, which is not really the case when you use AI to figure out what you want.
'I just needed to describe the spiral to something that would hold still long enough for me to hear myself.'
This line made me think of what a relief this kind of AI application can be for relationships. I have tried to explain this kind of spiral to my husband, a non-overthinker, and guess what? Not a good idea! Bless him, he really tries to understand, be helpful, make suggestions. None of it lands, though, because I didn't need a solution, I needed a sounding board. I have told him this, but he's just a "solutions kind of guy", which, at other times, is brilliant. ln the midst of the spiral, it just adds frustration, though. And Al doesn't mind if I get frustrated when it got me wrong (a. k. a. when I didn't have the language to say it right for it to understand at first).
The other thing I wanted to say is: definitely better than a rubber duck. Because the rubber duck doesn't ask point blank questions. Robot duck does.
And: I see you about how the sock brand matters. Same!
Haha oh how I hear you! My husband can't follow this thread either, I have a few friends that can but we spend an evening together and then we both need to take a break from each other for a month to recover.
The socks opener works, the rubber duck framing earns its place, and the input/output rule of thumb at the end is the kind of thing that actually sticks. You're writing from experience rather than performing a position, and the difference is audible throughout.
The Chi study lands cleanly, and the aside about handing your processed notes to dyslexic students not actually helping them — that's a quiet observation that carries more weight than you give it room for. Worth a longer look sometime.
One thing that didn't quite land for me: the "AI makes you stupid" paragraph. You either need to dive or gloss — what you do instead is half-dive, which invites the objection without really handling it. The frustrating part is that you actually have the right instrument for it. The generation effect is exactly the structural answer to that concern, not just a reframe of it. But you state it as a rule rather than argue it through, so it reads like assertion to anyone who isn't already nodding along.
And there's a harder edge just underneath it that goes unexamined: your rule of thumb only works if you're honest with yourself about which mode you're in. People outsourcing their thinking don't usually know they're doing it. That's where the argument gets genuinely interesting — and genuinely difficult — and that's precisely where the piece stops.
Which is a reasonable choice for a personal essay. But it means you're writing for the people who already agree. The rest of it is strong enough that it deserves a wider target.
That’s really useful feedback, thanks Peter! I know I’m writing for people who already agree, and I’m okay with that for now. I’m explaining, not arguing. But I appreciate knowing where the argument would need to go if I ever want to reach further.
I work a lot in the same way. Its all about discussion, and then my thoughts can crystalise into painting. I didn't need AI to be able to think, but I have found it so valuable in me organising my thoughts, that I no longer sit procrastinating for ages. Things snap into place so much quicker.
This is how I treat Claude Code most of the time. Meandering until my prompt actually resolves into a question (if not a resolution on its own). Pairs well with good context, but my prompting style has never conformed to the usual advice.
Very interesting. My wife and I were talking about this exactly today. And we do the same (with AI and with colleagues). And as you say this is something that we humans have been doing since before AI. The interesting thing is that I mostly come to my wife and sometimes she knows i just need to talk and sometimes she knows i need some feedback. But in many other occasions at work i tend to go to different colleagues depending on what i need or what im working on. I know you use it as just talk out loud method (and having a duck in the table would be good enough) but there is some type of interaction when we talk to a colleague or Ai and with ai we are always having the same (claude) answering entity. Would it be interesting having different AI personas or characters depending on what you need to discuss or what you need at the moment (just talk it out loud or actually some input here and there)? I guess that where the different points of view in combination with talking it out loud can be so powerful. Again i know you are focusing on just talking it out loud but its interesting why we might talk to certain people and not others (trust? Comfort? Familiarity?). Nice read
I loved this piece, this line stood out to me "The learning happened in the making. I was generating my own understanding, and the notes were just what was left over afterward." This is such an important concept, I'm just like you as I learn best after I put in the work. And I'm trying to instill the same learning into my kids.
Yes, wow. I feel you. Multiple ideas all threading & weaving & I also hardly look at what the output is when I am in it to win it. I will sometimes read it after I have gotten whatever intense thinking streak out of the way. & lately find more interesting parts in Claude thinking, not its output.
Sometimes I ask two different chats the same question and have multiple chats going at once.
The generation effect framing is what your piece does that mine didn't quite name. The reason articulating to AI works for AuDHD nervous systems isn't just that it answers back. It's that the act of producing language forces the loop to become linear. The brain encodes the act of producing, not just the result. For a monotropic mind running to resolution, that's exactly the cognitive shift that makes the burst close. I published an essay yesterday on a related angle (the journal that talks back, and why the talk therapy hour structurally can't hold AuDHD processing). Reading yours is the satisfying thing of seeing the same insight reached through completely different doors. Great piece.
I also think out loud and only recently realised conversations with AI help me understand how my thinking actually works rather than just organise it afterwards. It’s strangely reassuring to see someone describe the same process.
oh this is the same way my brain works. so many different threads, not reading the AI responses cos you're already ahead of it - especially chatgpt with its never ending rambling - but talking out loud to it and rubber ducking gives me so much more clarity. I've been trying to replicate it by doing audio recordings to myself, literally talking out loud, recorded just incase i say something interesting that i want to go back to one day. Makes me feel better knowing its not just my brain that thinks this way.
Love this, Emma! As I just said in my quote restack, I get why you related to "The Humans Who Think Like LLMs" now. I see this as a major element of constellation thinking. Sprawling, branching, and generative!
(Also I have NO idea why I didn't see this before. *gives Substack a sideeye* But I'm glad I finally did!)
I use AI strategically in this exact way. For routine, low-stakes decisions, I let AI handle the work. But for more critical thinking, I have AI adopt a role and interview me or push back on my ideas.
This approach was key during my recent brand pivot away from our legacy managed service provider role into business consulting. Having AI challenge my positioning forced me to articulate what actually differentiates advisory work from traditional IT services.
Your point about longer inputs meaning you're still thinking resonates. For higher sakes engagements , the real value of AI is not giving you an answer but helping you find the right question.
lol you basically described in this article my exact “method of prompting” the “act as this” alphabet soup prompting is useless, just define intent and if that changes half way through and the model doesn’t track the pivot then it’s time to get a better model. Love this article. I found you randomly but also one of my favorite follows is Prof. Sam Illingworth and his Slow Ai articles if you haven’t checked him out yet you should. I’m now Following you too to see what interesting things you find.
Thanks Michael! I think those prompting methods works well when you know what you want and wanna build something thats predictable, which is not really the case when you use AI to figure out what you want.
I think every one of my best prompts has started with the words “okay tell me if I’m crazy or if this is a thing?”
'I just needed to describe the spiral to something that would hold still long enough for me to hear myself.'
This line made me think of what a relief this kind of AI application can be for relationships. I have tried to explain this kind of spiral to my husband, a non-overthinker, and guess what? Not a good idea! Bless him, he really tries to understand, be helpful, make suggestions. None of it lands, though, because I didn't need a solution, I needed a sounding board. I have told him this, but he's just a "solutions kind of guy", which, at other times, is brilliant. ln the midst of the spiral, it just adds frustration, though. And Al doesn't mind if I get frustrated when it got me wrong (a. k. a. when I didn't have the language to say it right for it to understand at first).
The other thing I wanted to say is: definitely better than a rubber duck. Because the rubber duck doesn't ask point blank questions. Robot duck does.
And: I see you about how the sock brand matters. Same!
Haha oh how I hear you! My husband can't follow this thread either, I have a few friends that can but we spend an evening together and then we both need to take a break from each other for a month to recover.
The socks opener works, the rubber duck framing earns its place, and the input/output rule of thumb at the end is the kind of thing that actually sticks. You're writing from experience rather than performing a position, and the difference is audible throughout.
The Chi study lands cleanly, and the aside about handing your processed notes to dyslexic students not actually helping them — that's a quiet observation that carries more weight than you give it room for. Worth a longer look sometime.
One thing that didn't quite land for me: the "AI makes you stupid" paragraph. You either need to dive or gloss — what you do instead is half-dive, which invites the objection without really handling it. The frustrating part is that you actually have the right instrument for it. The generation effect is exactly the structural answer to that concern, not just a reframe of it. But you state it as a rule rather than argue it through, so it reads like assertion to anyone who isn't already nodding along.
And there's a harder edge just underneath it that goes unexamined: your rule of thumb only works if you're honest with yourself about which mode you're in. People outsourcing their thinking don't usually know they're doing it. That's where the argument gets genuinely interesting — and genuinely difficult — and that's precisely where the piece stops.
Which is a reasonable choice for a personal essay. But it means you're writing for the people who already agree. The rest of it is strong enough that it deserves a wider target.
That’s really useful feedback, thanks Peter! I know I’m writing for people who already agree, and I’m okay with that for now. I’m explaining, not arguing. But I appreciate knowing where the argument would need to go if I ever want to reach further.
I work a lot in the same way. Its all about discussion, and then my thoughts can crystalise into painting. I didn't need AI to be able to think, but I have found it so valuable in me organising my thoughts, that I no longer sit procrastinating for ages. Things snap into place so much quicker.
Yes there is something about that back and forth that really helps untangle our thoughts.
This is how I treat Claude Code most of the time. Meandering until my prompt actually resolves into a question (if not a resolution on its own). Pairs well with good context, but my prompting style has never conformed to the usual advice.
Very interesting. My wife and I were talking about this exactly today. And we do the same (with AI and with colleagues). And as you say this is something that we humans have been doing since before AI. The interesting thing is that I mostly come to my wife and sometimes she knows i just need to talk and sometimes she knows i need some feedback. But in many other occasions at work i tend to go to different colleagues depending on what i need or what im working on. I know you use it as just talk out loud method (and having a duck in the table would be good enough) but there is some type of interaction when we talk to a colleague or Ai and with ai we are always having the same (claude) answering entity. Would it be interesting having different AI personas or characters depending on what you need to discuss or what you need at the moment (just talk it out loud or actually some input here and there)? I guess that where the different points of view in combination with talking it out loud can be so powerful. Again i know you are focusing on just talking it out loud but its interesting why we might talk to certain people and not others (trust? Comfort? Familiarity?). Nice read
Loved this, especially the part about articulation changing your relationship to the problem instead of solving it.
Hi Emma,
You can read my short stories here:
https://gregleza.substack.com
Regards,
Greg
I loved this piece, this line stood out to me "The learning happened in the making. I was generating my own understanding, and the notes were just what was left over afterward." This is such an important concept, I'm just like you as I learn best after I put in the work. And I'm trying to instill the same learning into my kids.
Yes, wow. I feel you. Multiple ideas all threading & weaving & I also hardly look at what the output is when I am in it to win it. I will sometimes read it after I have gotten whatever intense thinking streak out of the way. & lately find more interesting parts in Claude thinking, not its output.
Sometimes I ask two different chats the same question and have multiple chats going at once.
The generation effect framing is what your piece does that mine didn't quite name. The reason articulating to AI works for AuDHD nervous systems isn't just that it answers back. It's that the act of producing language forces the loop to become linear. The brain encodes the act of producing, not just the result. For a monotropic mind running to resolution, that's exactly the cognitive shift that makes the burst close. I published an essay yesterday on a related angle (the journal that talks back, and why the talk therapy hour structurally can't hold AuDHD processing). Reading yours is the satisfying thing of seeing the same insight reached through completely different doors. Great piece.
I also think out loud and only recently realised conversations with AI help me understand how my thinking actually works rather than just organise it afterwards. It’s strangely reassuring to see someone describe the same process.
oh this is the same way my brain works. so many different threads, not reading the AI responses cos you're already ahead of it - especially chatgpt with its never ending rambling - but talking out loud to it and rubber ducking gives me so much more clarity. I've been trying to replicate it by doing audio recordings to myself, literally talking out loud, recorded just incase i say something interesting that i want to go back to one day. Makes me feel better knowing its not just my brain that thinks this way.
Love this, Emma! As I just said in my quote restack, I get why you related to "The Humans Who Think Like LLMs" now. I see this as a major element of constellation thinking. Sprawling, branching, and generative!
(Also I have NO idea why I didn't see this before. *gives Substack a sideeye* But I'm glad I finally did!)
I use AI strategically in this exact way. For routine, low-stakes decisions, I let AI handle the work. But for more critical thinking, I have AI adopt a role and interview me or push back on my ideas.
This approach was key during my recent brand pivot away from our legacy managed service provider role into business consulting. Having AI challenge my positioning forced me to articulate what actually differentiates advisory work from traditional IT services.
Your point about longer inputs meaning you're still thinking resonates. For higher sakes engagements , the real value of AI is not giving you an answer but helping you find the right question.