Big Idea

The argument in one line.

The reason AI learning sessions feel unproductive is that 'I don't know what I don't know' is a feeling, not a problem — and the only conversion tool is specificity, reached through a five-step repeatable method that ends with a prediction loop.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You use Claude Code and regularly stall because you don't know which option to pick or why an AI-generated plan looks right but feels opaque.
You're self-taught and worried you're pattern-matching responses without building genuine mental models that transfer to new problems.
You're building agent or sub-agent architectures in Claude Code and hit walls where AI answers feel plausible but you can't verify them.
You've noticed that AI research outputs age quickly and want a systematic way to date-check and verify claims before acting on them.

SKIP IF…

You already have a structured research and learning workflow — this is foundational, not advanced.
You're not using AI coding tools; every demo is Claude Code specific and some concepts won't map cleanly to other tools.

TL;DR

The full version, fast.

The phrase 'I don't know what I don't know' is a feeling, not a solvable problem — this five-step method converts it into one. Start by naming the gap as specifically as possible, then decompose it into sub-components to surface unknowns you didn't know you had. Run six accuracy checks on any AI answer: demand citations, verify them, triangulate across docs and forums and repos, anchor to the current year, probe for uncertainty, and ask what's missing. Reconstruct what you learned in your own words without looking at the source — if you reach for it, you faked a piece. Finally, predict, run a minimum viable test, compare, and treat every mismatch as the precise coordinate of your misunderstanding.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:54

01 · Why you don't know what you don't know

Hook question, the fuzzy-bubble-to-specificity concept introduced. 'I don't know what I don't know' is a feeling; 'I don't know X' is a problem.

00:54 – 03:28

02 · Step 1: Articulate the Gap

Vague vs. specific prompts. Live demo with superpowers brainstorm skill in Claude Code. The AI won't judge you — use that.

03:28 – 05:15

03 · Step 2: Decompose into Pieces

Custom illustrated slide: THE THING -> DECOMPOSE -> THE PIECES. Lists are tractable; fogs aren't. The X's are the real questions.

05:15 – 10:42

04 · Step 3: Explain and Verify

Six accuracy checks grid slide. Live research demo on Claude Code sub-agent architecture. Citations, triangulation, date-check 2026, probe uncertainty, ask what's missing.

10:42 – 12:18

05 · Step 4: Reconstruct

FAKE vs REAL illustration. Close the source and explain it in your own words. Language is the honest audit.

12:18 – 14:26

06 · Step 5: Apply and Test

Prediction loop diagram. Predict, run minimum viable test, compare, fork (mismatch = precise gift) or match (ship it). Full method recap slide.

Atomic Insights

Lines worth screenshotting.

'I don't know what I don't know' is a feeling. Feelings can't be solved. Problems can. The whole method is the conversion from one to the other.
Vague prompts produce vague answers because the AI has no target — specificity is the targeting system, not a courtesy.
Decomposing a black box into named pieces doesn't just help the AI answer better; it surfaces sub-questions you didn't know you had.
AI is a pattern-matching machine that confidently outputs whatever matches the pattern, true or not — six checks stop that.
Citations alone aren't enough: one article proves nothing. Triangulate across official docs, Reddit forums, and GitHub repos.
Date-check every AI research answer by explicitly asking for what's true as of 2026 — training data lags releases by months.
Asking 'are you sure you got all of it?' after a research session reliably surfaces new sources every single time.
Reconstructing knowledge in your own words without looking at the source is the only real test of understanding.
If you reach for the source to explain something, you faked a piece — language is the honest audit.
Mismatches in the prediction loop are precise gifts: they tell you exactly which piece of your mental model is broken.
A different AI model judging Claude's research output catches biases and gaps a single model systematically misses.
Persona prompting lets you probe uncertainty for free — ask Claude to take on an adversarial role and find what's missing.
The act of listing a concept's sub-components IS the work: the invisible gaps become visible the moment you try to enumerate the pieces.

Takeaway

Five steps from feeling lost to knowing exactly what to build.

WHAT TO LEARN

The gap between a productive AI session and a frustrating one is almost always a precision problem — and this method closes it in five repeatable steps.

Turning confusion into a solvable problem requires naming it precisely first — 'I don't understand agents' is not a problem, 'I don't know why an agent runs a planning step on every tool call' is.
Decomposing a large concept into its smallest sub-components is not a warm-up — the act of listing the pieces surfaces the invisible gaps that would have broken your build later.
AI research requires six active checks to be trustworthy: demanding sources, verifying them yourself, triangulating across source types, anchoring to the current year, probing for uncertainty, and asking what might be missing.
Reconstructing what you learned in your own words without referencing the source is the most reliable self-test: if you reach for the source while explaining, you faked a piece of understanding.
Mismatches between what you predicted and what the test returned are not failures — they are precise coordinates for the exact gap in your mental model, which collapses the search space for the next round.
Explicitly asking an AI model 'are you sure you got everything?' after a research session reliably surfaces new sources and caveats — use it every time before acting on research.
A second AI model reviewing the first model's output catches systematic biases you'd never find by re-reading the same source.

Glossary

Terms worth knowing.

Specificity Method: A five-step learning framework: articulate the gap, decompose into pieces, explain and verify, reconstruct, apply and test. Designed to convert a vague feeling of confusion into a precise, testable problem.
Skill (Claude Code): A reusable behavior or workflow installed into Claude Code that fires when its trigger conditions match a user prompt. The superpowers plugin uses a brainstorm skill that activates on 'I don't know what I don't know' inputs.
Sub-agent / fork: A Claude Code sub-agent is a child agent spun up to handle a specific task in isolation. A fork is the context-isolation mechanism that lets the sub-agent work with its own context window without polluting the parent's.
Triangulation: Using three or more distinct source types — official docs, user forums, and optionally a second AI model — to validate a claim, reducing the risk of trusting a single biased or outdated source.
Prediction loop: The test step of the Specificity Method: form a prediction, run the minimum viable test, compare results, and either exit (match = understood) or loop back to step 2 (mismatch = precise new gap identified).
Probing uncertainty: Explicitly asking an AI model after a research session whether it missed anything, which reliably triggers a second pass and surfaces additional sources or caveats.

Resources

Things they pointed at.

02:10toolSuperpowers plugin (brainstorm skill)

09:00toolNotebookLM ↗

00:00productAI Native community (Skool) ↗

Quotables

Lines you could clip.

00:48

“Specificity is the entire game.”

Punchy one-liner that stands alone with zero context→ TikTok hook↗ Tweet quote

04:10

“AI is a professional bullshit machine — it's just around to pattern match and pretend reason for everything that it does.”

Provocative framing that challenges common over-trust in AI outputs→ IG reel cold open↗ Tweet quote

10:54

“Language is brutal. You either have the concept or you don't.”

Hard truth about self-testing, no setup needed→ newsletter pull-quote↗ Tweet quote

12:21

“Mismatch is a precise gift.”

Reframes failure as signal — very shareable mindset flip→ TikTok hook↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphor

00:00You ever find yourself staring at Claude Code and thinking to yourself, I don't know what I don't know. How am I supposed to solve this? People tell me that phrase every single week, so I thought I would make a video to show you exactly how to deal with that problem.

00:11Okay. So our whole game over here is trying to get away from this fuzzy bubble of not knowing what you don't know to a point where you don't know a very specific thing. It's really that simple.

00:19We are trying to get to specificity here because when we have specificity and we can say, I don't know x, that means we have a very clear specific actionable problem that we can work on until we get to a point of understanding that matches a frame of reference we might have from something else out there.

00:35That could come through an analogy. It could come through a reference of something that you do understand in another field. However these things link, the important point is that you're only gonna get to these links for your mental model by being specific because specificity brings clarity.

00:49And we get to that specificity through these five simple processes over here that we're gonna run through now. The first one being, we need to articulate the gap between what we don't know and what we do know. And using a vague statement is definitely not the way to use that.

01:01For instance, if you say I don't get AI agents or this confuses me. Those are so vague that the AI does not know exactly how to help you. It might have a certain frame of reference of what to ask you next, but if you come this vague, you're going to get pretty vague answers up.

01:15If you had to pop over to Claude and say, I don't understand why this agent runs a planning step for every tool call, gonna get a much more specific answer. But in getting really specific over here, we're not just helping the AI. We're also helping ourselves understand what we don't know more clearly, which helps us build that mental model and also branch down different forms of questions that we can ask on top of that to get even more clarity and start to form an opinion on something.

01:37So how would we actually do that in practice? Now, the easiest way to do that is just to come over to Claude and tell it what you don't know. So you could say something like, I don't know what I don't know about AI agents.

01:47Can you help me get to a point of specificity? So the easiest way to use AI to become better is to do things like this. If you don't understand something, ask it to help you get to the point of clarity.

01:58The great thing about this is that it is not like being in front of someone really smart. Oftentimes, I find people might ask you a question and you give them an answer and they'll nod their heads as if they understand you, but they have that glossed over look in their eye where you know where they haven't really understood what you've been saying.

02:13The problem with that though is that you are not using AI to its full potential. This thing is not a human. It's not going to judge you.

02:18It will literally talk to you until it's explained the thing to you in a way that you will understand. So for me, I've got the superpowers plugin installed here. You absolutely don't need this.

02:26I wasn't intending for this demo to pull up this skill in the first place. It just matched that I don't know what I don't know about AI agents. So this skill matched on that, And now it's using its brainstorm functionality to go through helping me get to a point of specificity.

02:39And it does that by asking you follow-up questions about what exactly you don't understand. It might ask you what you're currently working on so that it gets context to then ask questions about that. And in this case, that's exactly what it's done.

02:50What's driving the need to get specific on agents? What would you do with the specificity once you had it? So this is trying to get context around what I'm building.

02:58So let's just say build agents for clients. We're gonna hit submit. It's gonna go in.

03:02We'll think and we just use this as a sparring partner to get to that point of specificity. I don't need to harp on about this forever because it's pretty self explanatory at this point. The reason why I brought this up is because this is obviously step one.

03:12This is how we get to our point of specificity. More importantly, it's where most people give up because they're so trained to not ask questions over and over again until they understand because of that whole judgmental thing that humans do to one another. Great.

03:24So we've articulated our problem. We've broken it down to a very specific thing. And the next part of this might be to decompose it into much smaller pieces.

03:32So if we come back to our environment over here, and we'd look at agents as a whole. So I just told this thing that I don't understand AI agents, blah blah blah, and it said good questions to ask. Agents is a huge territory and the right map depends on what you're trying to do with it.

03:45Let me start with the most important fork. So this is where we are decomposing that giant agents thing into small little avenues that we can explore. Because we said the AI that we want to sell this to clients, what exactly do we not understand about AI agents in this context?

03:58So it's offering us some smaller chunks that form part of this black box. It could be offer shaping and scoping behind selling an agent to a client. It could be the architecture involved in that.

04:07Maybe it's a little bit of sub agents as well and and how they work. But then inside there, what is a fork and how does a sub agent use that as opposed to the main context window? There are so many different avenues that you can explore once you start breaking this thing down into little chunks.

04:20And it's very important to do this because often even the first point of specificity that we have is not the most accurate one. What we're really saying is underneath layers and layers of other questions that form part of this bigger question that we're asking up top. So the point here where we're decomposing the big scary thing is by breaking it down into chunks, we are we are making the invisible things that we didn't even think of visible so that we can address those first as a part of our problem.

04:43Because I guarantee you, those of you who build with Claude and you don't know what you don't know, you're currently building, you might receive a plan and then you think, I don't know which one of these options to choose because I have no idea how several of the answers that it's giving me actually work. So that's why decomposing into these pieces helps you understand that framework.

04:59Once we get to that plan, you then know, oh, okay. I have to choose that because I know that sub agent's working this way and a fork makes sense over there because the main context window. Wherever your train of thought might lead you, the whole goal here is to understand how all of the pieces form the whole part, which makes your answer more confident because you understand the moving pieces on the chessboard.

05:17But before we can walk away and just accept everything that AI gives us, we need to understand that it is a professional bullshit machine because it is just around to pattern match and pretend reason for everything that it does. It doesn't think like humans do, which means we have to trust it less because pattern matching means as long as I met that definition of good over there and that kinda looks like what they were talking about, that means I did a good job.

05:37That's not true at all. So we fixed that using several things and you don't have to use every single one of these for every single problem that you have, but the more you use, the more accuracy you're going to get for your answers. The first thing here is to use citations.

05:48This is the most common one. It goes out there, it does research, and it puts the links of where it found these answers. The problem with that though is that just because some random dude online posted an article about something doesn't mean that he's right.

05:59So we can't just use one source. We have to use multiple sources to verify the citation, is part of step two. We're not just trusting some meta summary.

06:06We're not trusting a blog post. We're not even trusting maybe two websites. We're looking at a whole form of websites, maybe even mixing with other things over here.

06:14And that's where we triangulate our research. We take documents from the official websites. We might take some Reddit forums or some GitHub repos and see what people are talking about in there.

06:23We can dive into the software that we're looking at if it's publicly available in the GitHub repo. But by triangulating all this information, we get to see the user experience plus the official docs plus maybe even looking at some of the open source software. That gives us a much better opinion when you mix it with the citations, all the different forms of documentation.

06:40And then if you really wanted to, you could even use a different AI model to judge what this thing's output was based on all of these same things to see if the other model's biases found information that Claude didn't. But we're not finished yet because we still need to date check this thing. A lot of the videos out there now, you'll see when creators put things out within the first five minutes of something being released, you'll see that they are often missing gaps because they are using training data from months ago when an AI model was trained.

07:03And oftentimes, the problem that they're speaking about has already been solved and they don't even know it because they don't know AI and they're just using AI to go and do their research for them, but not properly. So for me, whenever I am doing research to validate a claim or learn something new, I make sure to tell the model check as of today in 2026.

07:19And this will go and not only just do all of my citations and verify across these various places that we spoke about, but it will also gather me the most recent information so that I know this thing's not just piggybacking off of old training data that it has. The next two things kind of tie in with one another. We have probing uncertainty and asking what's missing.

07:36I use these two all the time and I can guarantee you that if you do a round of research with Claude and then say, hey, are you sure you got all of the research right? Do one final sweep before before we move on to the next step. It will go.

07:46It will do more research. It will find new sources and give you new information. That's happened nearly every single time that I've done it.

07:51So using these two methods are really powerful, especially if you get a different AIM model to go and do this research for you. Of course, you don't really need to pay for another model to go and do this. You can use part of the free tier, or you can just use persona prompting where you can get Claude to take on a persona and go and probe for uncertainty or ask it what's missing from a different angle.

08:07I have a separate video that I'll put in the description on that below. So let's apply that research phase over here. I wanna know more specifically about architecture and framework choices.

08:15I'm talking specifically about Claude Code sub agents here. Let's take a look at the architecture behind that research, multiple sources here. I need citations.

08:23I want you to verify them. More importantly, I want you to look at Reddit forums and other forums where people have been building sub agents to validate any information that you find from the docs to ensure that that's the way that it actually works as a part of my business process. Now realistically, what you would have to do here is give a little bit of context about perhaps the skill or business process that you're trying to work.

08:43So let's get a little bit more specific about what we're trying to achieve here. And again, you would have done this with Claude to get to this point to then find out more information. I'm just using this as an example.

08:53I don't really understand how skill training works. Can you tell me a little bit more about the fork and how it saves context from the main context window when we're using sub agents? Also, know Anthropic just released something called forward slash workflows as a part of Opus 4.8.

09:08Can you tell me how that differs to skill training and sub agent forking or is it the same thing? So you can see here now I'm getting it to go and do all of this meta research, validate the sources, and I've also been really specific. Like I mentioned, you would have done this as a part of your fact gathering mission in these previous two steps that we've already done.

09:22Something else that I just wanted to point out on this research, what you can do is you can use NotebookLM. You can link it directly to Claude, have it pull down sources from YouTube or wherever it is that you generally get your information, and that will have your citations for you baked in there. Then you can just use Claw to go and pull those out if you want to.

09:38And so as our research starts to trickle back over here, we can start to sift through understanding whatever it is that the problem we're trying to solve. With our various references here, if you wanted to go and check to see that it was putting out in the correct places, I don't bother going that deep. I just trust that the information that it's putting forward from these multiple sources plus verification on top of that means I'm getting accurate results.

09:58That's also why at the end of the research, like I said, you're asking it again if it missed anything or if it's really sure about what it's putting forward to you. Sometimes I might even ask for a percentage and say, you a 100% sure that this information is accurate? And then it will go back and say, hey, I might have missed this.

10:12Point is it's still probably gonna be faster and better than any human in a given time frame. Your next part here is to understand how to apply this research to the problem that you're solving, which again if you still don't know how to do that part, you would just use AI to help you do that by saying, cool. I understand all of this information.

10:29How do I apply it to my situation that I'm trying to solve here? And then you guys would work together to get as specific as possible on the problem that you're solving with the new information that you've got. But the easiest way to understand if you're finally starting to get towards a clearer picture of understanding this thing is to take a test and see if you had to step back and talk to Claude.

10:48Could you have a discussion with it and say to it, oh, okay. So what you were really saying is that it works by doing x y zed. Basically, taking all of the information that you have now gathered from working with AI and putting it in your own words with your own explanations and your own terminologies and feeding that back into Claude so that it then either agrees with you or tells you no, your thinking is still wrong over here.

11:08The point is here is to reconstruct those thoughts into a mental model that makes sense for you. There's no point in taking AI jargon or jargon from other people's blogs and trying to recite that. It's much harder to remember other people's words in their own way.

11:20If you can turn something into your own framework with your own mental model and your own understanding of how it works within the solution that you're trying to put forward, it's way easier to remember and explain to another audience. So in this reconstruct part here, we would flip back into our chat and after I've gathered all this research, I might say, oh, skill chaining.

11:37Great. So skill chaining works by doing x y z and I know that I save all of these tokens because of a b c. Is that because it goes into the context window and does d e f?

11:47You get my point here. I always have this back and forth discussion. You can see that I've done it over here in some of my previous chats where I'm talking about data model mapping and I talk.

11:55I blatantly have this long conversation here using Whisper Flow and I will go through the thoughts how I understand it, what I think is the right thing to do. Generally, is a ton of swearing in my conversations because I'm usually pissed off by the time that I've researched over an hour of information. But the point is I'm trying to work to a point of clarity where not only do I get to the solution that I want, but to where I remember it in a way that makes sense for me and so that I can teach it to other people with my own opinion attached on top of it.

12:21And then finally, you made it to the end. All you have to do now is apply and test this thing. So we've worked through this with Claude.

12:27We've gotten to our specific questions and answers, and now we wanna go and test if the information that it's given us is actually accurate, and the solution that we can put forward on agents or architectures and what we're trying to do here is actually true. So we take that information, we've taken what we think we should do with our prediction loop over here.

12:44I think if I do x y zed, I'm going to get that result. Cool. Let's go and test this in a test run.

12:49So I'd come back over to Claude. We would go to our chat that we had open, which is over here. And I would just say to this thing, cool.

12:55Based on everything that we've done, let's go and test this. Can you build an environment to run an orchestrated test on this? And then Claude would say, sure.

13:01And it would spin up some sub agents. It would go and put them in different forks. It would be able to create the entire testing methodology for you in our example here, and then show you the results to figure out if all of the information that you guys have been discussing is actually true and accurate and works that way.

13:16If it doesn't, you're just going to learn. You're going to hit step two. You're going to run this thing.

13:20You're going to compare results and say, this doesn't really match what you said in the information, which means your research was probably a bit wrong or we didn't test it properly. In which case, it will go and iterate, check where it went wrong, do the loop again, and eventually, it will do this loop over and over again until you both have an understanding of how the thing works or you've achieved and built the thing that you actually wanted to build.

13:40And that is our five step framework in a nutshell. It's really not as complicated as it might seem. It's all about asking very specific questions, but before you even do that, making sure that you articulate the gap in your knowledge and you understand your problem at its smallest level.

13:54All of the little components that form part of that, then verifying the research and the information that you have. After you've researched it, then we reconstruct it, we put it in our heads so that we understand it in a way that makes sense for us. And then our last step here is just to apply and test everything because that way we are proving everything that we've learned.

14:10So I hope this video was helpful. I will link some skills that I use down below to help me troubleshoot these kinds of things and walk through getting to a place of specificity. If you have any questions, leave them in the comments below.

14:20I'll get back to you as soon as possible. Otherwise, check out the videos on the screen now. They'll definitely help you in your journey.

14:24Thanks very much for watching.

The Hook

The bait, then the rug-pull.

The most honest thing a builder can admit when working with AI is also the least useful: I don't know what I don't know. This video is built around the premise that feeling is not a problem — and the five-step Specificity Method is the conversion engine that turns it into one.

Frameworks