Big Idea

The argument in one line.

Default AI coding tools hand beginners a finished plan and a stack of 'recommended' choices to rubber-stamp, so a Claude Code plugin called VibeWise blocks code generation until the user has made the architecture and feature decisions themselves.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You're learning to code with Claude Code or a similar AI tool and want to actually understand the decisions behind your own projects, not just approve a plan.
You're a beginner or student who's noticed you click 'yes' through every AI recommendation without knowing what you agreed to.
You build side projects with AI and want a way to slow the tool down and force yourself into the design process.

SKIP IF…

You're a senior engineer who already understands the tradeoffs and wants AI to architect and execute end-to-end without friction.
You're optimizing purely for shipping speed on a project where learning isn't the point.

TL;DR

The full version, fast.

The bar for an impressive side project has risen because anyone can vibe-code one, but the classes and curricula teaching beginners haven't caught up, leaving no path to build the foundations AI quietly replaces. The creator built VibeWise, a Claude Code plugin that refuses to generate a full plan on request: it asks what stack to use, what the scope of version one is, how data should be modeled, and how edge cases like nested folders should behave, one question at a time, instead of offering a 'recommended' option to click through. The plugin gates code behind three checkpoint types, design, build, and implementation, so the user has specified and understood every decision before any line is written.

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 – 00:27

01 · The learning-versus-output trap

AI lets beginners ship more code, but the bar for an impressive side project has risen so much that simple AI-built apps no longer prove anything, while the classes meant to build foundational skill haven't kept pace.

00:28 – 01:25

02 · Introducing VibeWise

The creator built a Claude Code plugin called VibeWise that keeps the user in the loop by asking them to make architecture and feature decisions instead of generating them automatically.

01:26 – 02:54

03 · Demo: normal Claude Code behavior

A plain Claude Code session building a Notion clone offers recommended choices for scope and stack at every step, and the creator just clicks 'yes' through all of them despite saying he wants to learn.

02:54 – 04:13

04 · The fire-hose plan problem

Claude Code produces a full multi-thousand-line implementation plan and an 11-step build in seconds; no beginner is realistically going to read and absorb it, even though a senior engineer could use the same output productively.

04:13 – 07:29

05 · Demo: VibeWise's learning loop

Running the same 'notion clone' request through VibeWise, the plugin asks about programming experience, scope for version one, and then drills into open-ended behavior questions (sync across devices, block types, whether folders can nest) that force the user to actually spec the app.

07:29 – 08:23

06 · Checkpoints recap and close

The creator explains the design, build, and implementation checkpoint structure, code is only written after the design is fully specified, and closes by pointing to the GitHub repo and README.

Atomic Insights

Lines worth screenshotting.

The bar for a side project worth noticing has risen because to-do apps and simple CRUD tools can now be vibe-coded in minutes, erasing what used to distinguish a beginner's effort.
Classes have stayed roughly the same while AI capability has moved fast, so there's no longer a built-in path for beginners to develop the foundations AI now does for them.
Default AI coding assistants present choices as 'A (recommended), B, C' and most people simply click the recommended option and move on without engaging with the tradeoff.
A multi-thousand-line implementation plan generated in seconds is not something a junior engineer will actually read and absorb, no matter how clearly it's written.
Reverse-engineering a fully AI-generated codebase to 'learn from it' takes far more effort than most beginners will spend, so the learning rarely happens in practice.
For a senior engineer, a tool that plans and executes end-to-end is a genuine productivity win because they already understand the tradeoffs being made on their behalf.
Forcing an open-ended answer ('what should happen when you reopen the app tomorrow?') makes a learner actually think through a spec, where a multiple-choice prompt lets them guess and move on.
The small, 'obvious to an expert' decisions, like whether a folder can contain other folders, are exactly the decisions a beginner never learns to make if AI always fills them in silently.
Running into and solving unglamorous edge cases is what the creator frames as the actual fun of software engineering, not the typing of code itself.
Gating code generation behind design and build checkpoints means implementation only starts after scope, data model, and edge cases have already been decided by the user.

Takeaway

Multiple-choice AI defaults train you to stop deciding.

WHAT TO LEARN

When an AI coding tool offers a 'recommended' choice next to every decision, you'll click it every time, which means the decisions that used to build your judgment never get made by you at all.

A project being easy to vibe-code no longer proves anything about the builder's skill, since AI can produce a working to-do app or simple CRUD tool with almost no input.
Being offered a recommended option alongside every decision removes the friction that would otherwise force you to actually weigh the tradeoff yourself.
A long, auto-generated implementation plan feels like proof of thoroughness, but length alone doesn't mean a beginner will read, understand, or retain any of it.
Open-ended questions ('what happens when you reopen the app tomorrow?') force real thinking in a way that multiple-choice prompts with a recommended answer don't.
The 'boring' edge-case decisions, like whether a folder can contain other folders or what happens when a parent item is deleted, are exactly the decisions that build engineering judgment over time.
Gating code generation behind an explicit design-approval step is a workable pattern for anyone who wants to use AI as a Socratic partner instead of an autocomplete for entire projects.

Glossary

Terms worth knowing.

VibeWise: A free, open-source Claude Code plugin that blocks the AI from generating a full build plan immediately, instead asking the user a sequence of open-ended questions about stack, scope, and behavior before writing any code.
Vibe coding: Building an app by describing what you want and letting an AI model write and execute the implementation with minimal manual coding or review.
Build checkpoint: A pause point in VibeWise's flow where the user has to answer open-ended questions about what a feature should do before the plugin will continue.
Design checkpoint: A stage in VibeWise where multiple build-checkpoint decisions are combined into a single coherent feature design before implementation begins.
Implementation checkpoint: The final stage in VibeWise's flow, the only point where code is actually written, and only after the design has been fully specified and confirmed by the user.

Resources

Things they pointed at.

08:23toolVibeWise (GitHub repo) ↗

Quotables

Lines you could clip.

01:00

“I built a Claude Code plugin called VibeWise that basically forces you to be part of the loop and doesn't make you write code by hand, but leverages your mind to actually come up with a design.”

states the whole thesis of the video in one line→ TikTok hook↗ Tweet quote

03:20

“There's no way I'm going to be proactive enough to look into each of these options unless I'm really forcing myself to learn.”

honest admission that undercuts the default AI workflow→ IG reel cold open↗ Tweet quote

07:00

“That's what makes software engineering engineering, right? That's the fun part, or that should be the fun part.”

reframes edge-case debugging as the actual craft, not a chore→ newsletter pull-quote↗ 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.

Nowadays, people trying to learn how to code are kind of stuck between a rock and a hard place. They have to choose between using AI to increase their raw output, right? Increase their productivity, ship more lines of code, ship more features, and also prioritize their learning.

And so, for example, the bar for juniors has risen so much to the point where no longer is a, you know, to -do app or, you know, any sort of simple application going to cut it, right? Because those things can be vibe coded so easily and there's no way to distinguish. And so the bar for what is considered like a good side project has risen a lot.

But the problem with that is since the classes have stayed the same pretty much and AI has developed so quickly, there's not really a good way for them to build those foundations to be able to create these better applications without the use of AI heavily. And so I built a cloud code plugin called ViveWise that basically forces you to be part of the loop and doesn't make you write code by hand, but leverages your mind to actually come up with a design.

come up with the architecture, asks you questions to push you to, you know, come up with the features yourself, explain how you would implement the feature at a high level, right? So for example, it's not just going to create all the database tables for you. It's going to ask, what tables do you want?

And if you say, I don't even know what database to choose, it's going to ask you, which one do you think you should choose? It's going to keep pushing you to actually be involved in the decisions that you would have had to be involved in in the past and takes away kind of the multiple choice experience of A, best choice, B, C, which one you want to do.

And of course, everybody just clicks A and moves on. So without further ado, let's get into a demo. So the demo here is going to be pretty simple.

I'm just going to show an example of like a normal cloud behavior versus vibe wise. And so I created a new folder and initialized cloud. And all I said was I need to make a notion dupe.

That's literally all the context I gave. It asked me three questions. What is the clone for?

I said for my learning, which slice of notion should the first version cover? And then it gave me a recommendation to choose from block editor and pages. And then it said, which approach should we build on?

gave me a list of stacks to choose from, and then I just chose the recommended. So just look at this. I mean, I literally said that I wanted to learn, but then it offered me the recommended choice here, offered me the recommended stack here, and I just clicked yes, yes, yes, right?

There's no way I'm going to be proactive enough to look into each of these options unless I'm really forcing myself to learn, right? This is like the default behavior of AI to just kind of feed you everything, make it easy, so you can just get things done fast. And then now based on that, it says, Here's what you told me.

And I'm going to assume all of these things. I'm going to assume that it's single user, no login, no backend. I'm going to make a frontend only basically.

And then here's all the features I'm going to build. Here's the page tree, unlimited nesting, all of the different types of blocks, how it's going to search, how it's going to save. It comes up with the data model, the storage layer, the state, the routing, the editor, the styling.

I mean, think about how much it's doing here that you weren't even involved in. And it's just, you know, you're scanning it. It looks good.

You're like, okay, great. This is amazing. And then now all I have to do is say, yes, this looks good.

Does it look right? So I say, let's go. And then it literally comes up with a 11 step plan of what it's going to actually implement the, you know, project scaffold all the way down to how it's going to do each little feature.

And then now it's saying, you know, review the plan. Let's go check out what that plan looks like. You're telling me that like a junior engineer or somebody in college is actually going to scan this entire, how many lines is this?

You know, multi -thousand line file and actually understand, wow, I learned so much. You know, I came up with. all of this myself, I understand all the choices.

Of course not. They weren't involved in the building. They are not engineering anything.

They are saying what they want to be built and then AI is doing it from end to end, which is just, you know, could be great for implementing features, automated tests. For a senior engineer, this is amazing, right? Because they can architect everything and then actually understand what it's doing end to end without having to, you know, go through the learning process.

But for anybody trying to learn, this is a horrible experience. Because now you're overwhelmed because it basically gives you a fire hose of what it's going to do. And then how on earth are you going to learn from this?

You're telling me you're going to go read through this, you know, multi -thousand line document, understand everything, have it coded, and then read the, you know, thousands of lines of code that it coded and say you actually learned something? Of course not. Like, you know, of course reverse engineering is important, but there's no way you're reverse engineering this unless you spend a ridiculous amount of time and it's just...

Really not intuitive way to learn here. Versus VibeWise, which I can show an example of here. Once you install the plugin, you run it.

And I asked the same question. All I said was learn. And it said, what am I building?

And then I said, I need to make a notion dupe. And that's the same exact thing I said in the other, you know, terminal. But now it's saying, asking me questions.

How would you describe your overall programming experience? I'm a beginner. How do you want the learning loop to run?

You know, use default to ask questions as we go. And then it comes up with the progress, the profile, and the project map to keep track of it as we go. And then it asks you a build checkpoint.

What do you want the app to actually do when they open it, right? It's having you define the scope. It's having you think of, you know, what do I actually want my application to do when somebody clicks the URL and sees it for the first time?

And so I say, I want them to be able to open the app, sign in, create a note, and start editing the note. And you can see here, this is a free answer. Instead of giving you bullet point checklist or multiple choice in a question of what do you want to do with a little recommended next to it, it's having you actually think of an actual answer for what the scope is for version one.

And then it notes that down. And then next, there's another build checkpoint. We got to fill in the behavior.

There's so many gaps based on what you said. Signing in, creating a note, editing, that's great, but how do you actually want that done? And so it's asking you, what happens when you come back later?

How are you going to write a note? close the app, and then open it on another device. And then how's it going to be there tomorrow?

And then, you know, what does editing actually mean? Is it going to have these different types of blocks? So you can see here, it's asking about headings, you know, bullet list, checkboxes, having you come up with the actual spec.

And then how folders behave. It's asking you, well, you said you want folders. That's amazing.

It's not just going to fill in the gaps and, you know, come up with all the logic itself. It's saying, you know, can folders contain other folders? You know, only notes.

Maybe you couldn't note. exists outside of folders, what happens when it gets deleted, it's making you think about, you know, all of these little decisions that yes, are relatively simple. But if you've never experienced that, because all you're doing is just, you know, vibe coding, and basically reading a cheat sheet from AI, you'll never even have to think of these stupid little things that make up an actual application.

And by stupid, I don't mean that they're stupid, like decisions, it's just that, you know, to maybe a senior or somebody with a lot of experience, you know, I can look at this and just be like, yeah, of course, I understand in my mind without having to speckle this out exactly how I want it done. But if you've never had to think through the problem yourself, then you honestly probably wouldn't think of all these little edge cases and little behaviors, right?

That's the whole fun part of building is that as you build, you run into so many little things, so many little problems that you have to engineer. And that's what makes software engineering engineering, right? That's the fun part, or that should be the fun part.

If it's not fun for you, then you're probably in the wrong business. But this is what makes software engineering fun, is thinking about the user experience, thinking about how the backend works, how am I going to design the tables, what database do I even want to use?

There's so many little decisions that are just being abstracted away and being completely destroyed by AI when they honestly shouldn't be, especially when it comes to learning. And so, you know, give it a try. Let me know what you guys think.

But that pretty much sums up the demo. There's a lot of, you know, things in Readme that you can read through. Oh, the last thing I'll cover is that there's, you know, the different checkpoints, which is build checkpoints.

design checkpoints and implementation. And so implementation is the part where it actually writes the code, but only after you have been involved in the building process and finalizing the design, which kind of combines all the build checkpoints into one, which is really how you do features, right? You have to think of all the different things that make up an actual final design.

And then only once you have that design, do you look at it holistically, look at everything, how it works together, think of even more, you know, edge cases, more issues. And then only after all that has been flushed out and decided, then would you ever write the code? And so that's basically vibe coding, but actually having you be involved, right?

You're not actually having to write the lines of code yourself, but you are the one building it and designing it and actually having to think for once. So I'll put the link for this plugin in the description. You know, feel free to create any issues or just, you know, write feedback in the comments.

update this very, very easily. Thanks for watching.

The Hook

The bait, then the rug-pull.

Every beginner using AI to code is caught in the same trap: lean on the model and ship faster, or protect your own learning and fall behind the new, AI-raised bar for what counts as an impressive project. This creator's answer wasn't a tutorial, it was a Claude Code plugin that refuses to let you click 'yes' through someone else's decisions.

Frameworks

Named ideas worth stealing.

07:29model

VibeWise's three checkpoint types

Build checkpoint
Design checkpoint
Implementation checkpoint

Build checkpoints ask open-ended questions about one piece of behavior at a time; design checkpoints combine those answers into a coherent feature spec; implementation only fires once the design has been reviewed and confirmed, so code generation always comes last.

Steal forany workflow or plugin meant to slow an AI tool down enough for a user to actually absorb a decision instead of approving a default

CTA Breakdown