I also feel like there are possibilities for enhancing and extending the solution. For example, notebookLM has discover sources. Maybe that would speed up option one? Currently, it feels to me like this would take a while, but perhaps not as you suggest 30 minutes at the end.
Second, how to deal with domains where these frameworks don’t exist? Are high-quality sources still good enough for things to work without the strong support of frameworks?
Really great points, Paul. I think this system works for high-quality sources too (books, articles, courses, trainings, whatever materials) where the main step is applying them to your own situation. Imo, any learning material can function as a framework, or at least be transformed into one by thinking about how it becomes applicable.
And yes, NotebookLM could definitely accelerate Phase 1, and maybe even tools like ResearchRabbit, though that one leans more toward scientific work.
The time it takes really depends on how much effort you put into gathering and collecting the resources. Sometimes it’s long, sometimes not so long.
Hey Megan, I now use Claude the most. @Kamil Banc though create a really cool extension https://rcp.ad/ that helps you store your prompts easily. Hope this helps!
This is great! Would wrapping up all this process up into a custom GPT or an Artifact, or an app of some kind, be useful? Or is the human interaction per prompt the point.
It could definitely be turned into a custom GPT, I was actually thinking about that. I’d probably start from Phase 3 and 4 (application and evaluation), since those are more structured and repeatable. Phases 1 and 2 feel better when coordinated by a human, because mapping and validating expertise is a different type of process.
I love this!! This is a problem that badly needs to be solved, but almost no one even talks about tackling it, much less doing so. This is the first solution I’ve found.
Appreciate that, Paul! The gap between "AI sounds smart" and "is this actually good?" is pretty big indeed. Is there a specific area you’d want to try this system on?
Impressive. I'm gonna do this for SEO. If you did this for SEO because you use it, may want to do one for GEO. Search Perplexity or your favorite LLM for, Neil Patel and GEO. He's a no-shit expert.
Thanks, Ken! SEO was the test case I picked because it’s one of those areas where advice is everywhere but hard to evaluate. I haven’t dug into GEO yet, but I’ll check out Neil Patel. Thanks for the suggestion.
This is absolutely brilliant, Daria! Your four-phase system for strategic expertise borrowing perfectly addresses one of the biggest challenges in the AI era - distinguishing between AI's confident-sounding output and genuinely reliable expertise. I particularly love how you emphasized that "AI isn't replacing expertise, but helping you access and synthesize real expertise from people who've actually solved the problems."
The Harvard study findings you cited really drive home the transformative potential of this approach, especially for working outside our domains of expertise. Your detailed prompts for mapping credible sources, validating expertise, and creating systematic evaluation workflows are incredibly practical.
The key insight about using AI as a "knowledge engine" rather than just an "answer machine" is profound and will undoubtedly change how I approach complex projects going forward. The emphasis on that crucial Phase 4 evaluation step is spot-on - most people do skip this quality control step, which explains why so many AI-assisted outputs feel generic.
I'm curious about your experience with domains that are rapidly evolving - how do you handle situations where even the "experts" are still figuring things out in real-time, like in emerging areas of generative AI or cutting-edge biotechnology? Do you find the framework still holds, or do you adapt your approach for these fast-moving fields?
Thank you for the kind words, Shyam! I’m really glad you found it valuable, the best compliment to my work is when readers actually put these systems into practice 🙂
On fast-moving domains like biotech, it does get trickier. For those, I’d start with tools like Elicit, Scite, or Consensus (I covered them here: https://aiblewmymind.substack.com/p/5-ai-tools-that-do-the-research-for) since the most useful insights are usually in academic papers, not what Perplexity can surface. Scite, for example, lets you see where researchers agree or disagree, which helps you orient in the space.
I think the framework still applies if you curate strong sources, but what you do with them depends on your goal. And some baseline expertise is always needed to apply judgment once you move into actual execution.
Do you have a specific thing in mind where you’re thinking of applying this?
This is brilliant. You’ve taken what many people wish AI could do - magically fill their knowledge gaps and shown how to actually make that work in the real world. A proper system for pulling in proven frameworks and using AI as a thinking partner, not a substitute teacher.
According to Harvard Business School’s research, when people used AI in unfamiliar fields, their performance improved most dramatically - confirming everything you’re saying.
What’s been the hardest part for you - trusting the sources or building your own confidence around the output?
Thank you for the kind words, Melanie! tough question, I had to think about it before replying. I think for me the harder part is evaluating the output. You don’t really know if it’s “right” even if it cites sources.
The link to sources helps, but only if you actually stop to check them thoroughly, which is the step most of us are in a rush to skip. I remember once working with Perplexity on some data that perfectly confirmed what I believed. I built a whole thread of reasoning on it, and later checked where the data came from. Turned out the main source was a Reddit post.......
It’s so easy to fall into that trap because the speed these tools give us makes slowing down feel unnatural, but it’s exactly the slowing down that keeps the work grounded.
This is such a great piece — thank you for going this deep. I can see myself coming back to this again and again.
That last step of stress-testing it against the same standards was really eye-opening to me (not just because I might've been in the group that saw the original output as impressive enough, but also because..) I'm a little bemused that the LLM missed so much in the first place. Changing the prompt to "critique yourself" (vs "produce v 2.0") seems to be the key? I feel like even just knowing this will change how I use it in my daily work. Amazing piece, thanks again.
It really is the key, Karen! Any time you share a thought or opinion and ask AI to critique it or add counterpoints, it surfaces perspectives you might not have considered. And when you ask it to critique its own output, it does the same for its reasoning. This little trick really does wonders.
I also feel like there are possibilities for enhancing and extending the solution. For example, notebookLM has discover sources. Maybe that would speed up option one? Currently, it feels to me like this would take a while, but perhaps not as you suggest 30 minutes at the end.
Second, how to deal with domains where these frameworks don’t exist? Are high-quality sources still good enough for things to work without the strong support of frameworks?
Really great points, Paul. I think this system works for high-quality sources too (books, articles, courses, trainings, whatever materials) where the main step is applying them to your own situation. Imo, any learning material can function as a framework, or at least be transformed into one by thinking about how it becomes applicable.
And yes, NotebookLM could definitely accelerate Phase 1, and maybe even tools like ResearchRabbit, though that one leans more toward scientific work.
The time it takes really depends on how much effort you put into gathering and collecting the resources. Sometimes it’s long, sometimes not so long.
I agree, the important part is to apply to your own situation. That works best in connecting dots and internalizing knowledge.
Really appreciate the article Daria! As always, full of great and practical insights!
Thank you, Jenny! Yes, in the end, it all comes down to application so that it becomes useful.
Daria, I always read your posts twice because you pack in so much value! Thanks for sharing Guidde, I’m going to play with it 🤗
That makes me so happy to hear, Karo! Guidde has been a great find for me too 🤗
Love how you make AI so easy to understand always educating.
Thank you, Sae! You made my day with this 🙏
Great piece! Where do you typically store and organize all your prompts? Do you find this is most effective in Claude or ChatGPT?
Hey Megan, I now use Claude the most. @Kamil Banc though create a really cool extension https://rcp.ad/ that helps you store your prompts easily. Hope this helps!
This is great! Would wrapping up all this process up into a custom GPT or an Artifact, or an app of some kind, be useful? Or is the human interaction per prompt the point.
It could definitely be turned into a custom GPT, I was actually thinking about that. I’d probably start from Phase 3 and 4 (application and evaluation), since those are more structured and repeatable. Phases 1 and 2 feel better when coordinated by a human, because mapping and validating expertise is a different type of process.
That seems great! Seems like the perfect blend.
I love this!! This is a problem that badly needs to be solved, but almost no one even talks about tackling it, much less doing so. This is the first solution I’ve found.
Appreciate that, Paul! The gap between "AI sounds smart" and "is this actually good?" is pretty big indeed. Is there a specific area you’d want to try this system on?
Impressive. I'm gonna do this for SEO. If you did this for SEO because you use it, may want to do one for GEO. Search Perplexity or your favorite LLM for, Neil Patel and GEO. He's a no-shit expert.
Thanks, Ken! SEO was the test case I picked because it’s one of those areas where advice is everywhere but hard to evaluate. I haven’t dug into GEO yet, but I’ll check out Neil Patel. Thanks for the suggestion.
Great post Daria!
Thank you, Neil!
This is absolutely brilliant, Daria! Your four-phase system for strategic expertise borrowing perfectly addresses one of the biggest challenges in the AI era - distinguishing between AI's confident-sounding output and genuinely reliable expertise. I particularly love how you emphasized that "AI isn't replacing expertise, but helping you access and synthesize real expertise from people who've actually solved the problems."
The Harvard study findings you cited really drive home the transformative potential of this approach, especially for working outside our domains of expertise. Your detailed prompts for mapping credible sources, validating expertise, and creating systematic evaluation workflows are incredibly practical.
The key insight about using AI as a "knowledge engine" rather than just an "answer machine" is profound and will undoubtedly change how I approach complex projects going forward. The emphasis on that crucial Phase 4 evaluation step is spot-on - most people do skip this quality control step, which explains why so many AI-assisted outputs feel generic.
I'm curious about your experience with domains that are rapidly evolving - how do you handle situations where even the "experts" are still figuring things out in real-time, like in emerging areas of generative AI or cutting-edge biotechnology? Do you find the framework still holds, or do you adapt your approach for these fast-moving fields?
Thank you for the kind words, Shyam! I’m really glad you found it valuable, the best compliment to my work is when readers actually put these systems into practice 🙂
On fast-moving domains like biotech, it does get trickier. For those, I’d start with tools like Elicit, Scite, or Consensus (I covered them here: https://aiblewmymind.substack.com/p/5-ai-tools-that-do-the-research-for) since the most useful insights are usually in academic papers, not what Perplexity can surface. Scite, for example, lets you see where researchers agree or disagree, which helps you orient in the space.
I think the framework still applies if you curate strong sources, but what you do with them depends on your goal. And some baseline expertise is always needed to apply judgment once you move into actual execution.
Do you have a specific thing in mind where you’re thinking of applying this?
Thanks for sharing this 😁 really great piece that I've Restacked for my audience - they'll love it ✨
Thanks so much, Chris, appreciate the restack!
Interesting read. Thanks for sharing your thoughts on this.
Glad it's useful!
This is brilliant. You’ve taken what many people wish AI could do - magically fill their knowledge gaps and shown how to actually make that work in the real world. A proper system for pulling in proven frameworks and using AI as a thinking partner, not a substitute teacher.
According to Harvard Business School’s research, when people used AI in unfamiliar fields, their performance improved most dramatically - confirming everything you’re saying.
What’s been the hardest part for you - trusting the sources or building your own confidence around the output?
Thank you for the kind words, Melanie! tough question, I had to think about it before replying. I think for me the harder part is evaluating the output. You don’t really know if it’s “right” even if it cites sources.
The link to sources helps, but only if you actually stop to check them thoroughly, which is the step most of us are in a rush to skip. I remember once working with Perplexity on some data that perfectly confirmed what I believed. I built a whole thread of reasoning on it, and later checked where the data came from. Turned out the main source was a Reddit post.......
It’s so easy to fall into that trap because the speed these tools give us makes slowing down feel unnatural, but it’s exactly the slowing down that keeps the work grounded.
This is such a great piece — thank you for going this deep. I can see myself coming back to this again and again.
That last step of stress-testing it against the same standards was really eye-opening to me (not just because I might've been in the group that saw the original output as impressive enough, but also because..) I'm a little bemused that the LLM missed so much in the first place. Changing the prompt to "critique yourself" (vs "produce v 2.0") seems to be the key? I feel like even just knowing this will change how I use it in my daily work. Amazing piece, thanks again.
It really is the key, Karen! Any time you share a thought or opinion and ask AI to critique it or add counterpoints, it surfaces perspectives you might not have considered. And when you ask it to critique its own output, it does the same for its reasoning. This little trick really does wonders.
Thank you for the kind words!
Wow, Daria. Just wow.
Bookmarking this one📌
Glad it's useful, Bechem! Can't wait to hear how it worked for you and what use case you used it for
Tedt