A tier-list breakdown of ten popular pieces of AI advice, ranked S to F by what actually compounds versus what quietly wastes time or money.
Posted
3 days ago
Duration
Format
Listicle
educational
Views
9K
127 likes
57 · 43
Big Idea
The argument in one line.
Most popular AI advice fails because it optimizes the wrong variable: cheap models trade money for worse output, and the habits that actually compound are skills, context discipline, and planning before building.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You're using AI daily to build a product or run a business and want to stop guessing which habits actually compound.
You already have a Claude or ChatGPT workflow but keep hitting token limits, context bloat, or scattered tool subscriptions.
You manage multiple AI agents or sub-agent tasks and want a working model for splitting work without losing track of it.
SKIP IF…
You're brand new to AI tools and haven't built a repeatable workflow yet, most of this assumes an existing practice to audit.
You're looking for step-by-step prompt engineering tactics, this is about systems and habits, not specific prompts.
TL;DR
The full version, fast.
Ten pieces of common AI advice get ranked from S to F based on what actually compounds versus what quietly wastes time or money. The worst advice chases cheap open-source models to save cents while losing hours, tries to build everything instead of buying proven tools, and chases every new AI tool that launches. The best advice turns repeatable work into reusable skills, builds a structured personal knowledge base AI can actually search instead of relying on one-off chat context, plans before building instead of improvising, and has AI verify its own output against a stated plan before anything ships. The throughline: own your workflow instead of renting intelligence one conversation at a time.
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.
Austin introduces the S-to-F tier board ranking ten pieces of popular AI advice and previews the format.
00:35 – 01:59
02 · Strategy 1 (D-Tier): Switch to free open-source models
Argues that chasing free models trades money for wasted time or worse output, and reframes AI spend as a cost of doing business rather than a waste.
01:59 – 03:15
03 · Strategy 2 (A-Tier): Turn repeatable workflows into skills
Explains why one-off AI conversations don't transfer knowledge anywhere, and lays out a four-step method for turning a completed task into a reusable skill file.
03:15 – 04:09
04 · Strategy 3 (C-Tier): Give AI as much context as possible
Context helps only up to a point; stale or conflicting reference files quietly degrade output, so audit what's feeding the AI periodically.
04:09 – 05:21
05 · Strategy 4 (S-Tier): Build an LLM knowledge base
Describes a personal knowledge base split into a raw folder and an indexed wiki layer, letting AI retrieve high-quality context on demand instead of drowning in files.
05:21 – 07:26
06 · Sponsor break: Whop
Mid-roll integration for payments platform Whop, demoed through a real screen recording of setting up product tiers from inside Claude Code.
07:26 – 08:51
07 · Strategy 5 (F-Tier): If you can build it, build it
Warns against rebuilding tools you already pay for; introduces the 10X-vs-2X filter for deciding whether a build is worth the time.
08:51 – 10:09
08 · Strategy 6 (A-Tier): Plan before you build
Makes the case for Claude's plan mode, with the caveat to actually scrutinize the plan rather than rubber-stamp it, and not to over-plan past the point of useful information.
10:09 – 12:00
09 · Strategy 7 (F-Tier): Automate everything
Lays out a five-step process (question, delete, simplify, accelerate, then automate) borrowed from Elon Musk's production discipline, warning that automating a broken process just scales the breakage.
12:00 – 12:28
10 · Subscribe ask and Claude Max giveaway
Anti-slop disclosure, subscriber goal, and this episode's Claude Max subscription giveaway winner.
12:28 – 14:10
11 · Strategy 8 (B-Tier): Use multiple agents to build faster
Three tactics for running agents in parallel without losing track: launch independent sub-agents for split research, separate deep work from small surface tasks, and give each agent a non-overlapping task.
14:10 – 15:29
12 · Strategy 9 (D-Tier): Experiment with new AI tools
Argues against constantly trying new tools due to decision fatigue and relearning costs; recommends waiting for a tool to prove itself before adopting it.
15:29 – 16:59
13 · Strategy 10 (S-Tier): Give AI a way to check its own work
Self-verification roughly doubles output quality but is generic by default, so the fix is explicitly building a verification step into the plan or asking AI how to manually verify the result.
16:59 – 17:32
14 · Final tier list and close
Recaps all ten strategies on the completed S-to-F board and points to a follow-up video on the knowledge-base system.
Atomic Insights
Lines worth screenshotting.
Open-source AI models can be free in dollars but expensive in time, and time is the cost that compounds fastest.
A $500 car payment isn't wasted money if the car earns you more than $500, the same logic applies to AI subscription spend.
Teaching an AI assistant how you work inside one chat thread is a rental, the knowledge disappears the moment that conversation ends.
A reusable AI skill is built in four steps: do the task, ask AI to turn the conversation into a skill, run it, then correct and re-save it.
More context isn't automatically better, outdated or conflicting files in an AI's context window hurt output quality more than missing context does.
A personal knowledge base works best split into two folders: raw source material, and a wiki layer that indexes what's actually useful in it.
Rebuilding software you already pay $10 a month for rarely pays off once you count the hours and the ongoing maintenance burden.
The real question before building anything isn't 'can I build this', it's 'is this a 10X move or a 2X move'.
Planning only works if you actually scrutinize the plan, nodding along in plan mode defeats its entire purpose.
Action produces information: overplanning can be as costly as underplanning because some requirements only surface once you start building.
A five-step process before automating anything: question the requirement, delete what you can, simplify what's left, speed it up, then automate.
Automating a broken process doesn't fix it, it just produces garbage at scale faster.
Running multiple AI agents works only when each one owns a distinct, non-overlapping task, otherwise context-switching erases the time saved.
Chasing every new AI tool that launches creates decision fatigue, wait for a tool to prove itself before adopting it.
Having AI check its own work roughly doubles or triples output quality, but its default verification is still generic until you define what 'correct' means for your specific task.
Takeaway
Ten AI habits ranked by what actually compounds
WHAT TO LEARN
The habits that compound, skills, a structured knowledge base, planning, and self-checking AI, trade short-term effort for long-term leverage, while the habits that feel productive usually cost more time than they save.
02Strategy 1 (D-Tier): Switch to free open-source models
Open-source models can look cheaper per token, but the time burned working around their limitations usually costs more than the subscription you skipped.
Reframe AI spend like a tool you need for work, not a discretionary expense, the token budget is the cost of building at all.
Current AI pricing is subsidized by investor money, so build the discipline to use tokens efficiently now before prices rise.
03Strategy 2 (A-Tier): Turn repeatable workflows into skills
A one-off conversation with AI teaches it how you work, but that knowledge disappears the moment the chat ends, it never transfers anywhere else.
Packaging a workflow into a portable skill file means it moves with you across tools instead of staying locked to one chat thread.
Build skills task by task: do the work once, ask AI to turn that conversation into a skill, run it, then correct and re-save it.
04Strategy 3 (C-Tier): Give AI as much context as possible
More context helps only up to a point, past that it becomes noise AI has to dig through to find what actually matters.
Outdated or conflicting reference files hurt output quality more than missing context does, audit your files periodically and remove what's gone stale.
05Strategy 4 (S-Tier): Build an LLM knowledge base
A personal knowledge base works best split into two layers: raw unprocessed material, and a wiki layer that indexes the key takeaways from it.
Structuring your own knowledge this way lets AI find high-quality context fast instead of searching through everything you've ever written.
The goal isn't collecting more information, it's building a map so AI can retrieve the right piece of it on demand.
07Strategy 5 (F-Tier): If you can build it, build it
Before rebuilding a tool you already pay for, count the real cost: the build time, the missing edge cases, and the ongoing maintenance.
Ask whether a build is a 10X move or a 2X move, 2X moves usually aren't worth the time even when they're technically possible.
08Strategy 6 (A-Tier): Plan before you build
The best builders move slow to move fast: decide whether to build something first, then figure out how, rather than diving straight in.
Plan mode only works if you actually scrutinize the plan AI proposes, nodding along defeats the entire point of planning.
Don't over-plan either, some requirements only surface once you're actually building and testing, action produces information planning can't.
09Strategy 7 (F-Tier): Automate everything
Automating a broken process just produces garbage at scale faster, fix the process before you automate it.
Run five steps before automating anything: question the requirement, delete what you can, simplify what's left, speed it up, then automate last.
Be ruthless about deleting steps you're not sure you need, if nobody complains once it's gone, you didn't need it.
11Strategy 8 (B-Tier): Use multiple agents to build faster
Parallel agents only save time when each one owns a distinct task, overlapping work creates context-switching costs that erase the speed gain.
Split independent research across several agents at once rather than running it sequentially, that alone can multiply throughput several times over.
Separate deep, high-focus work from small surface tasks, batch the small stuff across several parallel threads instead of doing it one at a time.
12Strategy 9 (D-Tier): Experiment with new AI tools
Trying every new AI tool that launches creates constant decision fatigue and relearning costs that outweigh whatever marginal edge the new tool offers.
You don't need to be first to adopt a tool, if it's genuinely useful it will keep showing up until it's worth trying.
13Strategy 10 (S-Tier): Give AI a way to check its own work
Having AI verify its own output before you see it can roughly double or triple the quality of what it hands back.
Default self-verification is generic, AI doesn't know what 'correct' means for your specific task until you define it inside the plan.
Ask AI how you can manually verify a result after any big change, it will often hand you a usable verification checklist.
Glossary
Terms worth knowing.
LLM knowledge base
A structured set of files on your own computer, split into raw source material and an indexed summary layer, that an AI assistant can search to answer questions about your business or work.
Task-driven skill creation
A four-step method for turning one completed AI task into a reusable skill: do the task, ask AI to extract a skill from it, test the skill, then refine it based on results.
10X vs 2X
A filter for deciding whether to build something: a 10X move changes a project's trajectory enough to justify the effort, a 2X move is a marginal improvement not worth the build time.
“The whole point of planning is to use your brain, so you have to think.”
blunt, punchy, works alone→ TikTok hook↗ Tweet quote
16:30
“AI doesn't know what correct looks like for your specific task.”
clear limitation statement, sets up the actionable fix→ IG reel cold open↗ 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.
17px
metaphoranalogystory
In this video, I'm going to break down the best and worst advice for using AI to build 10 times faster. Because the truth is, there have never been more people who are sharing horrible AI advice that is completely backwards, and it is destroying how people use AI in their day to day. Now, if you don't know me, my name is Austin, and I'm not just a content creator.
I was a lead engineer at JP Morgan, I was a COO of a tech startup worth over $25 million, and I've helped hundreds of founders and operators dominate with AI in their specific industry. So for each AI strategy, I'll break it down in a tier list, then break down exactly what should be doing differently instead and the nuances that are only clear after you've spent hundreds of hours using these tools so strategy number one which is d -tier is reduce your token spend by switching to free open source models this advice is well intentioned and i know from the comments that budget is one of the biggest concerns for a lot of you but let me explain why this is a problem there are three ways you can actually pay for ai with your money with your time and with the quality of your work open source models may save you money but you'll pay for it somewhere else either in the time you burn or in a worse output and this is exactly why i tell everyone only use free open source models when you have to and since you're watching this video you're likely one of two people you're there's someone using ai to build a business or you're someone using ai to help you at your full -time job and neither of those is something to treat lightly and i'm not saying to go crazy and burn through tokens what i am saying is rewire how you think about ai spend in the first place if you need a car for work you don't call your 500 car payment a waste of money you need it so you spend it and your ai spend is
that same logic. If hitting token limits or chasing cheaper models is hurting your output, the fix isn't cutting a tool. It's rethinking the spend.
And the reality of the situation is right now your tokens are subsidized by VC money. The $100 you're spending today could cost $2 ,000 once these companies stop chasing market share and start chasing profit. So what should you do instead?
Well, no matter what, make sure your system is token optimized so you get the most out of your budget. And if you're wondering how to do that, link below. I have a free walkthrough that uses my Claude plugin.
you do exactly this. So that's the first strategy. Getting to strategy number two, which is A tier.
Turn repeatable workflows into skills. There's a foundational belief that I've talked about a lot on this channel. When you're using AI, you want to own the intelligence, not rent it.
And skills are one of the best ways to unlock that. If that sounds a little bit complicated, let me explain. When you go back and forth with Claude, Chachi Petit, or whatever tool, you're teaching that system how you work.
But there's a problem with this. Even if it nailed exactly how you work, that teaching doesn't go anywhere. It lives in that one conversation on that specific platform.
And that dynamic of being platform specific, that isn't owning intelligent. That's a rental and it puts you at the mercy of whatever the AI service provider decides to charge you next. This is where skills come in.
It packages up your industry knowledge into a file that you own. So if you build on Claude, you could bring it into Codex and it'll work the same way. Now, skills are super powerful and I won't dive into it too much in this video, but there are some nuance when using this.
You don't want to go around and collect every skill you find on the internet like you're collecting Pokemon. Instead, use tech. task -driven skill creation, which has four steps.
First, do the task, then tell AI to create a skill from that conversation, then run the skill and see if it works. And then at the end, if it doesn't work correctly, correct the output and then say, based on this conversation, enhance the skill. And as you get into creating skills, you're likely going to add more and more direction to them.
And that brings us to the next piece of advice. Strategy number three, which is C tier, is give AI as much context as possible. So to understand this, you need to understand what context is actually.
This is the information about you your business whatever that you give ai to help you produce a better output this information is hyper critical but people don't always understand the nuance more high quality context is good but only up to a certain point the example i like to use is think of the context you give ai like a fridge ai opens the fridge sees one high quality item and knows exactly what to grab you add a bit more it's fine but keep stuffing the fridge and eventually it's impossible to find anything in there and then if something goes bad within the fridge the whole fridge can smell and it can impact everything else.
So there's a rule of thumb that I go by. Relevant context helps. Outdated or conflicting context hurts.
So to audit your own setup you can use this prompt to review every contextual file you've got and flag anything that's gone stale. Now if you're wondering what the actual best way to provide context is that's where strategy four comes in which is S tier. Build an LLM knowledge base your AI can actually use.
This is one of my favorite pieces of advice on this entire list because it genuinely changed the trajectory of my career. An LLM knowledge base is a way to structure contextual files on your computer to better teach AI about your job or your business. Going back to the fridge metaphor, what we're building here is an industrial refrigerator room.
This allows you to add more context and it's still easy for AI to find it. The easiest way to understand this is to just show you mine. So I call it my internal OS and inside it, I have two folders, my raw and wiki.
Slash raw is where every unprocessed piece of information sits. So for example, this could be the exact article. I liked about writing marketing copy.
And then in the slash wiki folder, it takes the material and then pulls out the key learnings. It's essentially creating a table of contents for everything that's sitting in the slash raw folder. Then when I ask the knowledge base question, it'll check the wiki first.
And when it needs more information, it knows exactly which raw file to look for. You're essentially building a map for AI so it can find and access high quality contextual information for whatever you're working on. Now I have a deep dive video about this if you want to go further, or if you want me to help you set this up for your business, there's a link below to where with me either way once ai understands more about your work you'll start seeing more things you can build which makes the next mistake especially tempting which is f tier but before we get to that everything we're talking about is about building faster but what if you're ready to take the next step and start getting paying customers which brings us to today's video sponsor wap which i'm super excited about if you've ever wanted to turn something you've built into a real business the idea usually isn't the hard part it's really everything around it that's taking payments getting new customers and knowing what's actually working this is what wap handles which i've been really
impressed with. So you guys know that I've been working with buildpartner .ai, my Claude plugin that helps people build faster with Claude. And WAP now officially powers the payments for it.
And yes, previously I did have a payment processor, but why did I move? The reason is that WAP isn't just a way to process payments. It's everything that's built on top of the payment layer.
Now there are three features that I love. The first is that you can run your business from inside Claude code. So WAP has a CLI and an MCP that can connect directly to Claude.
So you can just ask it in plain English to make a payment link, pull your stats or launch an ad this is actually how i set up all of buildpartner's payment tiers without ever leaving cloud code on screen this is literally the prompt that i wrote i want you to create all of my current products in wop and it just did it the second feature is the flexible checkout so customers get 100 plus payment methods without any additional setup plus they have a buy now pay later so a customer can pay over time while you get paid entirely up front the third is that it can help you with ads this is something that i'm personally excited to get my hands dirty with because i'm not an ad expert.
I primarily do organic content like this, and I bet a lot of you watching aren't ads experts either. But the reality is ads are one of the best ways to get traffic to your product, and WAP has made it super easy to get them set up and running. And because their ad platform is tied to real payment data, you can see which ad actually made a sale, not just which one got the clicks.
So whether you're starting something new or already have a business, sign up for WAP using the link below in the description. Now, no matter what tool you use for payments, you have to be strategic with what you build.
Which brings us to the next piece of advice, which is F tier. Strategy five, F tier. If you can build it, you should build it.
So AI has made it possible to build almost anything, but should you? So sure, you could rebuild the software you're paying 10 bucks a month for, but how long will that actually take you? And will it hold up through every edge case?
And are you going to do the maintenance when it needs a patch every six months? This makes it obvious. The answer is no.
And I've watched so many people learn the basics of vibe coding or see how powerful something like log code really is. And then suddenly they want to build absolutely everything. So here's my question for you.
Do you actually need to build what you want to build or are you just And I've personally fallen for this, right? I once decided to rebuild our CRM myself.
About 10 hours later, I just never used it. It didn't hit feature parity. And I just decided to go with...
Google Sheets. And I've watched business owners do this time and time again. They've rebuilt an entire project management tool over a single missing feature only to realize that they were misusing the original one and they wasted all of that time.
Building on this, I recently read a book called 10X is easier than 2X. And the core idea is that 10Xing is actually easier than 2Xing because it kills the decision fatigue of debating whether to do something at all. So before you build anything, ask yourself, is this a 10X move or is it a 2X move?
And then this will make it extremely clear clear that when you want to rebuild a software that you already have or it just is missing one single feature you just won't do it because that's a 2x move. And if you're watching this channel we're 10xers we're not 2xers.
Save the 2xers for other channels. Now you will eventually decide to build something and that brings us to the next piece of advice. Strategy six which is a tier plan before you build.
This is building on the last point but everyone in the AI world wants to go go go but the best people actually move slow to move fast. So the first question you should always ask is should I build this?
But when the answer is yes, it becomes how should I build this? And that's where plan mode is so powerful. In Claude, just type slash plan and it drops you straight into plan mode.
Or you could just say interview me to create a plan on how to build whatever you're trying to build. But with this, there are a couple of nuances worth thinking about. The first is that you have to scrutinize the plan.
Everyone falls into this trap, but if you're going to just nod along, you might as well not be in plan mode. The whole point of planning is to use your brain so you have to think. The second is is that don't overplan.
My favorite quote on building fast comes from Brian Armstrong, the CEO of Coinbase, and he says, action produces information. So yes, you should plan, but don't plan so much that you never actually build anything. As I've built automations and products, I've noticed that you will never think of everything during the planning phase.
Some requirements really only surface once you're actually testing and using the thing that you want to build. So plan, but don't overplan and bias towards action. That's how you get the information.
And once you start planning and building, you're going to... get the building bug. And at that point, you're going to want to follow this next piece of advice.
Strategy seven, which is F tier, automate everything. Now this one is different from the build everything trap. It's not about if you should automate something, it's about how you should automate something.
Automating your whole life sounds sexy on paper, but it skips four steps that you need to hit first before you do anything with automation. And this actually is inspired by Elon Musk. Say what you want about the guy, but he runs a five -step process before auto anything.
The first step is question every requirement. Why do this? Why do I care in the first place?
Make requirements less dumb and never assume a rule is correct just because a smart person made it. Step two is delete any part of the process that you can. You want to be ruthless about removing things and if you're not occasionally adding things back, you're probably not deleting enough.
This is something I started doing more and more in my business where if I'm not sure if information is needed, I just delete it and then if somebody complains, we'll figure it out later. And I have yet to have anybody complain when we delete processes. So that's step two.
Now, step three is simplify and optimize. Improve what's left after deletion. Elon Musk warns that the most common error of a smart engineer or operator is optimizing something that shouldn't exist.
Step four is accelerate cycle time. Before you automate anything, can you just speed up whatever you're doing before actually automating it? This is the process of automating versus augmenting.
Can you augment it enough so that there's no longer friction in actually completing the task? And then step five. automation which always comes last automating a broken process just produces garbage at scale and the first four steps make sure that that doesn't happen and now before we get to the last three strategies one of which includes my favorite you guys know the drill this is our anti -slop agreement everything here was built by humans and for humans so as part of this agreement subscribe to this content so that i can keep making it and it can reach more and more people we're at 90k subs and i have a goal of getting to 100k subs in the next month so i appreciate everybody who subscribed And as a thank you, I give away a Claude Max subscription every video.
And this video's winner is Chase Ferrari 2291, who is building a tool to promote personal brands. To enter this video's Claude Max giveaway, comment the worst AI advice you've ever personally been given or the best advice you've gotten. I will accept sarcastic answers as well.
Let's have some fun in the comments. Now, once you've found something that's worth automating, if you split work across agents, that can be extremely helpful, but it only works if it actually separates, which brings us to strategy. Strategy eight, which is B tier, use multiple agents to build faster.
This is genuinely strong advice, but there's one nuance you have to get right. The more agents you're running, the more context switching you have to do. Each time you jump from one task to another, you lose time remembering where you were.
And we've all experienced this, right? You jump between tasks and you feel like you never got anything actually done. Now, I'm not saying you should wait there for five minutes while AI works on something, twiddling your fingers, but you should be strategic about how you use agents.
Here are three strategies. that have actually worked for me in terms of paralyzing tasks. The first is say launch sub -agents in your prompt if you have tasks that could be split into independent pieces.
For example, I might say launch five sub -agents to research across YouTube, Google, Reddit, Twitter, and my internal OS. That paralyzes the research instead of running it sequentially, which makes it roughly five times faster. The second is partition deep work and surface work.
If something's genuinely important and takes all your focus, just do that one thing at a time. But if it's all small tasks, that just need to get done, that's the perfect candidate for running four or five threads at once.
My personal move here is just spinning up four or five separate chats in cloud code, one per task, and archiving each one once I'm complete with that specific task. The third is give each agent a unique task so that you can avoid overlap. The best way to think about this is how I run buildpartner .ai.
I've got one agent working on admin features, another on the landing page, and another on our one -on -one coaching feature. Each owns a distinct part of the product, so I'm not worried about it crossing swords. Now Zoom Zooming out, those are specific pieces of advice, but this is a general trap that most people fall into.
This is the trap of more, and this doesn't just stop at agents. For example, tools. More tools isn't automatically better either, which brings us to strategy nine, which is D tier, experiment with new AI tools.
So this is a weird one, but I hate this piece of advice for a couple of reasons. The first is that a new tool drops every single day. The moment you allow yourself to start trying them, you now have to decide which ones are worth your time and which are worth skipping.
And this is just stressful and it introduces unnecessary decision fatigue. The second is if you do start trying new tools, you end up constantly switching between tools. And every switch means reconfiguring and relearning something you already know how to use.
Now, personally, I have one AI workspace, which is Claude. Then I have two or three other tools that solve a very specific use case or problem I have. So here are two foundational rules that help most people that I talk to.
The first is look for tools that solve specific problems. Don't window shop just for the sake of it. The second is that you don't need to be first to try a tool.
You're on this channel, so you're naturally in the AI ecosystem. If a tool is actually worth it, it'll keep popping up and you'll be able to use it. Don't worry.
If you take a look at my channel, I intentionally don't try and be the first person to talk about tools because I actually practice what I'm preaching here. I don't care to be first. I don't care to be second.
I do care to adopt it. And once I feel it's needed, I will make a video about it. Now, whatever tool you choose, making sure that something actually works and isn't broken is...
super critical to getting high quality outputs from AI. Which brings us to strategy 10, which is S tier. Give AI away to check its own work.
This is something that the creator of Claude Code said will 2 to 3x the quality of your outputs. And it is so powerful that Claude's most recent models now do this by default. So what does this actually mean?
And if the model does it by default, do you still have to care about this? Well, first, what does it actually mean? If AI can proactively catch issues in what it created and then fix them before you ever see it, it's like they never happened.
But there is a limitation to this though. verification is still somewhat generic. AI doesn't know what correct looks like for your specific task.
So to minimize needing to go back and forth, you still have to strategically build the verification element yourself. And here's how I personally do this. For more technical tasks, I'll say, tell me how you'll verify the output as part of the plan.
Now, I've already talked about how powerful plan mode is, but by explicitly building in a verification plan, you'll be more confident that AI actually knows what the target finished product is. Then for non -technical tasks, I define the output structure I want ahead of time. script outline skill and as part of that it verifies the final structure before I ever see it.
Both of these at a high level will get AI verifying its own work but what I personally like even more is having AI tell me how I can verify the output. After any big change I'll say how can I manually verify this and it hands me a personal verification plan. The reality is that most people still verify outputs by hand and they should but they neglect to have AI help them with this actual step in the verification process.
So you want to think about how you can use AI to make your life easier in your manual verification. So let's run through the board.
D through F tier, that's chasing free models, drowning in context, building things you don't need to, and chasing every new tool that drops. This is all terrible advice. The top tiers, that's turning repeatable works into skills, planning before you build, running agents strategically, building a knowledge base AI can use, and having AI check its own work.
Just by doubling down on the winners and avoiding the bad advice will transform your productivity. Now one A tier piece of advice was about building llm knowledge basis so if you want to learn how to do that go watch this video if you like this i guarantee you'll love that so i'll see you over there peace
The Hook
The bait, then the rug-pull.
Austin Marchese stacks ten pieces of popular AI advice on an S-to-F tier board and works through each one in turn, from the "switch to free models" line that quietly costs more than it saves to the knowledge-base habit he calls the one that changed his career.
Frameworks
Named ideas worth stealing.
02:55list
Four Steps of Task-Driven Skill Creation
Do the task
Tell AI to create a skill from that conversation
Run the skill and see if it works
If it doesn't work correctly, correct the output and say 'based on this conversation, enhance the skill'
A method for building reusable AI skills from real completed work instead of collecting pre-made skills.
Steal forbuilding a personal skills library instead of grabbing every public skill off the internet
04:09model
LLM Knowledge Base (raw / wiki split)
/raw — every unprocessed piece of information (articles, transcripts, source material)
/wiki — pulled-out key learnings, acts as a table of contents pointing back into /raw
A two-folder structure that lets AI find high-quality context fast without drowning in every file ever saved.
Steal fororganizing a personal or business knowledge base so AI tools can actually use it
10:09list
Five-Step Process Before Automating Anything
Question every requirement
Delete any part of the process you can
Simplify and optimize what's left
Accelerate cycle time
Automate — always last
Borrowed from Elon Musk's production discipline; automating a broken process just produces garbage at scale.
Steal forauditing any repeatable business process before building automation around it
12:59list
Three Strategies for Parallel Agent Work
Launch independent sub-agents for tasks that split cleanly (e.g. parallel research across platforms)
Partition deep, high-focus work from small surface tasks and batch the surface tasks
Give each agent a unique, non-overlapping task so ownership never crosses
A model for running multiple AI agents without losing the time savings to context-switching.
Steal forcoordinating multiple Claude Code sessions or sub-agents on one project
14:56list
Two Rules for Evaluating New AI Tools
Look for tools that solve a specific problem, don't window-shop for the sake of it
You don't need to be first, a genuinely useful tool will keep showing up
A filter against tool-hopping fatigue.
Steal fordeciding whether to adopt a new AI tool the moment it launches
CTA Breakdown
How they asked for the click.
VERBAL ASK
05:21product
“sign up for WAP using the link below in the description”
Native mid-roll folded into the S-tier knowledge-base segment: shows a real screen recording of the sponsor's dashboard and his own product setup inside Claude Code before making the ask, rather than a cold interruption.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
A JP Morgan engineer turned $25M startup COO lays out the seven pieces of Claude he'd learn again, in the exact order that compounds instead of the order most people stumble into.
A six-step roadmap for turning Claude Code into a personal operating system — build breadth across domains first, then compound depth in the one vertical you already work in.
00:13
00:30
00:59
01:12
01:25
01:38
01:46
02:13
02:27
02:43
02:57
03:16
03:33
03:51
03:56
04:29
04:38
04:55
05:03
05:35
05:45
06:01
06:14
06:30
06:41
06:55
07:09
07:15
07:49
08:01
08:13
08:26
08:39
09:03
09:18
09:37
09:42
09:58
10:18
10:37
10:50
11:04
11:17
11:30
11:43
12:09
12:35
12:44
13:02
13:15
13:33
13:47
13:59
14:14
14:34
14:47
15:00
15:20
15:39
15:53
16:11
16:13
16:30
16:45
17:12
17:23
21:11
Video of the Day18:06
13:10
19:04
23:27
14:05