Claude Code's Creator Reveals Every Level of AI (and How to Level Up)
A viral Boris Cherny tweet mapped five stages of Claude Code adoption — this breakdown turns the jargon into a pizzeria anyone can picture.
Posted
1 weeks ago
Duration
Format
Essay
educational
Views
5.1K
142 likes
57 · 43
Big Idea
The argument in one line.
A viral adoption ladder sorts AI-coding-agent users into five stages, from chatting with an AI that never touches real work to running thousands of autonomous agents, and moving up a stage always means breaking a specific bottleneck and building a specific guardrail, not just using more tokens.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You lead or work inside a team already using Claude Code and want a plain-English way to tell how far your adoption actually is.
You're comfortable with Claude Code day to day but haven't set up auto mode, worktree isolation, loops, or channel-monitoring agents yet.
You want a mental model for what changes — bottleneck and guardrail — at each step of trusting an AI agent with more autonomy.
SKIP IF…
You've never used Claude Code or any AI coding agent — the pizzeria analogy assumes you already know what an 'agent' does.
You're looking for a hands-on setup tutorial with exact commands and config — this stays at the conceptual/framework level.
TL;DR
The full version, fast.
A viral tweet from Claude Code's creator mapped five stages of AI adoption, and this video translates the jargon using a pizzeria analogy: at Stage 0 you're the only cook and AI just shouts advice through the mail slot; by Stage 4 you own a thousand kitchens that hire, restock, and open new locations without you. Each stage swaps a specific bottleneck (permissions, your attention, review load, trust in the loop, finding enough work) for a specific guardrail (tool access, spend caps, auto mode, written standards, cost controls). Roughly a third of users are stuck at Stage 0 and only a tenth of a percent reach Stage 4.
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.
The host previews Boris Cherny's viral tweet outlining the levels of Claude Code adoption and promises to translate the technical table into plain English.
01:00 – 03:05
02 · What Boris tweeted about ascending
Reads Boris Cherny's original thread: tokens alone don't move you forward, each step has its own bottleneck and guardrails, and Anthropic itself sits at level three pushing toward level four.
03:05 – 04:47
03 · Level 0: The Tadpole
Introduces the pizzeria analogy — AI advice shouted through a mail slot, no agent has touched real work yet. About a third of users are stuck here; the bottleneck is permissions, not technology.
04:47 – 06:47
04 · Level 1: Assisted, The Supervisor
One agent is let inside and given tool access, but every move must be watched before it ships. Roughly 45% of users sit here; the bottleneck is the supervisor's own attention.
06:47 – 08:45
05 · Level 2: Parallel, The Head Chef
Five to ten agents run at once with auto mode on and self-checked work; the role shifts from supervisor to head chef reviewing diffs. About 20% of users reach this stage; weeks of backlog compress into an afternoon.
08:45 – 12:39
06 · Level 3: Supervised Autonomy
Dozens to hundreds of agents run continuously using loops, batch, and worktree isolation; spot-checks replace per-task review. Only about 2% of users are here; the bottleneck is trust in the loop.
12:39 – 15:14
07 · Level 4: AI-Native, God Mode
Thousands of agents operate across the business via an agent SDK and channel-monitoring tools that kick off work proactively. Under 0.1% of users reach this stage; the bottleneck shifts to finding enough work worth automating.
15:14 – 16:16
08 · How to set this up yourself
The host pitches his paid AI community for personalized, step-by-step help implementing the levels covered in the video.
Atomic Insights
Lines worth screenshotting.
Adoption of an AI coding agent breaks into five stages, and roughly a third of users never get past the first one, where AI gives advice but never touches real work.
The bottleneck at the earliest adoption stage is permissions, not technology — the AI is capable, it just hasn't been given access to any tools.
At the supervisor stage, one agent works but every change still needs human review before it ships, which caps output at whatever one person can inspect.
Moving from supervising one agent to orchestrating five to ten of them changes the job from checking keystrokes to reviewing final diffs, and a week of backlog can compress into an afternoon.
Only about two percent of users reach the stage where agents run continuously in the background without per-task review — the bottleneck there is trust in the loop, not compute.
At the highest stage, under a tenth of a percent of users run thousands of agents that start their own work; the constraint becomes finding enough well-scoped work to hand them, not running more agents.
Isolating each parallel agent in its own git worktree keeps simultaneous agents from overwriting each other's changes.
A written standards file plus reusable skills is what lets autonomous agents encode team conventions instead of needing a human to explain them every time.
Turning on auto mode removes the need to approve each individual action, which is the guardrail that unlocks running multiple agents at once.
Using a flat-rate subscription instead of pay-per-token API access removes the cost anxiety that otherwise limits how many agents someone is willing to run.
The advice at every stage is the same shape: break the current bottleneck, then build the guardrails the next stage requires, rather than trying to skip ahead.
Takeaway
What Actually Changes At Each AI-Adoption Stage
AI ADOPTION LADDER
Moving up the five adoption stages means breaking one specific bottleneck and building one specific guardrail at a time, not simply feeding the AI more tokens or trusting it blindly.
02What Boris tweeted about ascending
Tokens alone don't move you to the next adoption stage — you have to find and break that stage's bottleneck and set up its guardrails.
Even Anthropic, the company behind Claude, only sits at stage three of its own ladder and is still working toward stage four.
03Level 0: The Tadpole
At the earliest stage, roughly a third of users are only chatting with AI for advice — it never gets permission to touch their actual work.
The blocker at this stage is permissions and access, not the AI's capability, so the fix is granting tool access, not switching models.
04Level 1: Assisted, The Supervisor
One agent gets access to your tools, but you still have to watch and review nearly every change before it ships.
The bottleneck shifts from access to your own attention span — you can't step away while the agent works.
05Level 2: Parallel, The Head Chef
Running five to ten agents at once with auto mode on turns a week of backlog into an afternoon of orchestration.
Your job changes from checking every keystroke to reviewing final diffs, which is what makes running multiple agents in parallel sustainable.
06Level 3: Supervised Autonomy
Only about two percent of users reach the stage where agents run continuously in the background without a per-task review.
At this stage you fix the recipe — the underlying instructions or context — rather than fixing individual outputs one at a time.
Isolating each agent in its own git worktree is what keeps a large number of parallel agents from overwriting each other's work.
07Level 4: AI-Native, God Mode
Under a tenth of a percent of users reach the stage where thousands of agents start their own work without being asked.
The constraint at the top stage isn't running more agents, it's identifying and scoping enough legitimate work to hand them.
Glossary
Terms worth knowing.
Auto mode
A permission setting that lets an AI agent execute edits and commands without asking for approval on every single action.
Worktree isolation
Running each parallel agent in its own separate git working directory so simultaneous agents can't overwrite each other's changes.
MCP governance
Organizational rules controlling which external tools and plugins (via the Model Context Protocol) an AI agent is allowed to connect to and use.
Agent SDK
A programmatic toolkit for building and scheduling AI agents in code, rather than driving them one prompt at a time in a chat interface.
Loops / batch / routines
Features that let an agent run repetitive or scheduled work automatically on its own trigger, instead of a person kicking off each task by hand.
“Personally, Boris just hit level four, which is good for Boris, but I personally just hit level six.”
cocky escalation punchline right after naming the ceiling→ TikTok hook↗ Tweet quote
03:24
“The world's best chef stands outside your pizza shop shouting recipe tips through the mail slot.”
vivid single-image analogy that explains the whole first stage→ IG reel cold open↗ Tweet quote
04:24
“The bottleneck here is the permissions, not the technology.”
tight reframe that resets the viewer's assumption about what's holding them back→ newsletter pull-quote↗ Tweet quote
07:42
“You review final diffs rather than keystrokes, and the backlog of maintenance starts shrinking.”
concrete before/after of what changes at the parallel stage→ newsletter pull-quote↗ Tweet quote
13:24
“Your phone only rings, or you are only needed, not to approve work, but if something goes wrong.”
crisp definition of what 'god mode' actually feels like day to day→ 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.
17px
metaphor
Boris Cherny, the creator of Claude Code, just revealed the four distinct levels of Claude Code users and how we can go from beginner or intermediate to an expert operating on god mode. And this tweet went viral getting 1,400,000 views.
He says, I talk to engineers at other companies every day and hear the same thing. One person is 10 x ing their output with Claude, but the rest of the organization hasn't caught up.
Boris watches teams adopt AI, and he keeps seeing the same four steps or levels of adoption. He mapped them out here in the steps of AI adoption. And when you click on the link, it pulls up this magnificent table where Boris talks about going from level zero all the way to level four of being a Claude code user and exactly what you need to do to transcend these steps.
Now this table, honestly, if you read it, it is a lot of technical jargon. And so in this video, I wanna make it as simple as possible for us to understand exactly what he's talking about so we can all ascend to god mode Claude code users together. Now before I jump into my simplified version of Boris' table, I want to show you exactly what Boris tweeted with this link because he gives us some hints on how to actually move up the ladder in these tweets.
He says, there's no one right path through the steps. Every team and company is different, but at each step, tokens aren't enough to move you forward. To get to the next step, you need to find and break down the next set of bottlenecks, which we're gonna go through.
Every step has its bottlenecks, and every step has its guardrails that allows you to actually operate on that level. So you have to set up the guardrails for the next step, and you have to break the bottlenecks of this step, and then you ascend.
Boris said in practice, that means giving Claude ways to verify its own work end to end so you can remove the human from the loop. It means enabling auto mode for permissions, defaulting on automated code review, security review, and using interfaces that let you manage multiple agents at once, which is very important.
Even talking about called tag here, which we're gonna get to when we talk about god mode. To get higher levels, it means using loop, batch, dynamic workflows, and work tree isolation for sub agents.
And at a glance, this is some technical stuff, but we're gonna break it down together. It's not about a single feature, but rather using the right features with the right guardrails that enable Claude to automate entire classes of work in a way that your team can trust the output so humans aren't even in the loop.
I'm excited to show you exactly what this god mode setup looks like because, honestly, it's pretty mind blowing. Boris then goes on to talk about how do you track the extra productivity by ascending the levels, which isn't too relevant for us, but he finished by saying the bigger payoff comes when fixing and maintaining happens in the background and your teams or yourself can just focus on building.
That's when you start doing things that weren't even in range before. Anthropic, as a company, he says, is on step three or level three of ClaudeCode usage and pushing towards level four, which is the god mode setup. Personally, Boris just hit level four, which is good for Boris, but I personally just hit level six.
So we'll get into that in a sec. Boris says, curious where you are. What step is your team or what step is your business or you personally on?
Well, to figure that out, we come over to this simplified version of the steps. And starting with the tadpole step stage zero. And you can see that there's about 33% of Claude users are actually stuck on step zero.
What does this look like? And we're gonna be using a fun example in this video together of us running a pizzeria. So if we were to run a pizzeria at Stage 0, it would look like this.
The world's best chef stands outside your pizza shop shouting recipe tips through the mail slot. Of course, the world's best pizza chef is the AI, and we're in there doing the work. There are no chefs inside.
There's no AI actually doing the work. They never touch the food. You cook every single pizza yourself, but you can literally just talk to the mail slot and ask for its advice and then go and do the work.
Remember, this is level zero. What your role is, you are the cook. Now down here, I've put the exact table from Boris' original post, which means you have zero agents.
You're using AI, but you're not using agents. Agents are the version of AI who can actually go and do the work on your behalf. What it looks like, Boris, is a lot of people are using older or lighter or faster models.
Latency compounds through AI gateways. There's no MCP governance, so no plug ins to actual tools and softwares that you're using. Internal access to AI tools is gated or process heavy.
No IT infrastructure or approvals for hosting Claude created code or artifacts, so every output exists only locally. In plain English, you chat with AI, you're chatting with it through the letter slot in the door, but it does not touch your work.
It does not touch the pizzas. So the bottleneck here is the permissions, not the technology.
You have not given it permission to touch the software and tools inside your computer. Products that help, Boris says claude.ai is just what people are using as a chat, and the guardrails you need to set up are these ones right here, which come pretty default.
But then we ascend to level one, which is the assisted level where you become the supervisor.
So let's start with the pizza analogy. The chef is now finally allowed in, off the street, out of the cold. One chef, one kitchen station.
But you stand right at their elbow watching every stretch of the dough, checking every topping before it lands faster than cooking alone because you have a chef who's cooking for you and the chef is more skilled, so they can remember where this is an AI analogy. It's not just a chef analogy. They can put the pepperoni on.
They can chop faster. They can dice the tomatoes faster just like your AI can write an email faster than you can. But the catch here, and this is the bottleneck, you can never step away from the counter.
You are a supervisor, and you can't stop watching AI work. But you do finally have an agent because you have your Claude code plugged into your software and your tools, and now they can operate them on your behalf without having to ask permission. It's the only thing that'll send you from stage zero to stage one.
Boris says it looks like one engineer, one agent, mostly supervised, a fast pair programmer. You run one session at a time and review almost every change before it merges with your live products. So whether you're creating a website, you've got a software, or you've got a store, an ecommerce store, before you publish it, you need to check through everything.
The bottleneck is your attention and the need to inspect each response and code edit or output edit if you're not a coder specifically. At stage one, you've gone from using Chord inside the browser to using Chord code in the desktop, CLI, or IDE, which is basically download the desktop app or maybe use Chord code inside your terminal window, which is how you can give it permission to actually use your laptop, your hardware, and your software on your behalf.
And the guardrails is basically if you are using API, which I highly recommend you don't, you basically got to put in the how much money it's allowed to spend on each task and, of course, a centrally managed policy that's in line with your business. But we don't wanna stay supervisor forever. So how do we ascend to stage two or level two, which is called, according to Boris, parallel.
I call it the head chef, which there are 20% of people at this stage, which is a pretty awesome stage to be at, especially in the early days of AI. Now in our pizza analogy, we now have 10 chefs at 10 stations, and they're all cooking at once.
I've stopped cooking entirely because I'm now head chef, and I walk the line tasting each pizza before it goes out the door. My job has changed from supervisor to head chef.
You can see it. We now have 10 agents working for us, 10 chefs in the line. And Boris says what it looks like is one engineer or you orchestrating five to 10 agents at once, each on its own work tree or git checkout jumping between them.
Claude checks its own work, tests, builds, security scans before you see it. Auto mode is always on. You do not need to keep giving it permissions.
It has permission to cook the pizza. It has permission to make the website to change features inside your software app. Boris says your output multiplies, which is why it's very important to get to stage two if you're not here already.
You review final diffs rather than keystrokes, and the backlog of maintenance starts shrinking. The unlock here, Boris says, is a backlog that used to take the team weeks of work, guys. Weeks of work now become your afternoon of orchestrating your 10 agents to do it on your behalf.
The bottleneck, of course, is you reviewing it before it goes live, and the products that help is turning on auto mode inside your Chord, agent view, Chord code review, Chord security review, using your Chord code on your mobile device while it actually executes in the cloud on your desktop, starting to use Chord tag, which we're going to come back to because that is a massive breakthrough.
And, of course, using remote control so you can monitor your agents from your phone and keep them working if they finish a task and given the next one. Now if you are running a team, you need analytics to monitor the team usage, and you need Claude powered end to end verification to make sure you're setting up all its access to everything that Claude needs in order to keep you out of this loop, of course, apart from approving its work.
So how do you go from Headchef to level three where only 2% of Claude users are currently at? And Boris calls this supervised autonomy.
I'm calling it the manager of head chefs. You can see you're just ascending basically the org chart of a typical business.
In our pizza restaurant, for example, we now have hundreds of chefs and then tens of chefs as head chefs managing those chefs. Bit of chef ception for you here.
Now there's too many pizzas to taste because we have hundreds of chefs cooking them, but we do spot checks. And when one comes out wrong, we fix the recipe card. We fix the formula going in, not that output, not that pizza directly.
We are now the manager of the multiple head chefs who manage those 100 chefs beneath it. And this is why sometimes you'll see clips of Boris where he's like, yeah.
I've got a thousand agents running right now, and you're like, how and why? And what are they even doing? This is why.
Because he was pioneering level three of being a Claude code user where you do have a 100 agents. Boris says, Claude writes all or nearly all of the code. Instead of asking it, did you read the code?
It becomes what context was the model missing and how do we solve it for next time? You're thinking in batch operations. Remember?
A recipe card, not the actual specific outputs. What you unlock at this level. As I've started to breach level three for my business, I can see this just frees up a lot of time and puts you in a crazy exponential mode that your competitors just won't be able to keep up with you.
The unlock here is clawed proactively does work that you would have had to kick off manually before, so you would have had to be there to start the actual work and to watch over it. But maintenance and cleanup that used to wait for someone to find the time now runs continuously in the background.
AI does nearly all the work and starts chores by itself without you even needing to be there. This is the loops without review. The bottleneck at this stage, Boris says, is trust in the loop that you set up.
And if you haven't watched our loops video by Boris where we go through the four types of loops that he is actually running and how to set them up for yourself, doing literally walk through examples for each of them, I'll link it up in the description or in the top for you to go and watch that. So the bottleneck, he says, is trust in the loop and your team's decision throughput.
The agent tree is too deep to babysit, and your trap is scaling agent count before the loop has earned widespread trust. You also wanna make sure that you're using tokens efficiently if you're not just using the subscription, which I highly recommend. And, of course, how we actually are running this is using sub agents with work tree isolation so the parallel agents don't collide and override each other's work.
Running routines, running loop. Once again, go watch that Boris video if you want more information there.
Running batch and running goal to fan out and schedule repetitive work. Doing dynamic workflows, which is actually a very fascinating concept that Boris has linked up, which I'm gonna be doing a video on, so make sure you subscribe if you wanna see more Boris content in the future. And now Claude Tag to have it monitor a channel or data source and kick off tasks proactively.
So this is the new feature, the omnipresent setup where Claude is inside of your Slack channel. Now I actually don't use Slack, but I've been using it on Telegram.
You can also set up on WhatsApp for yourself so that Claude is reading all of your team messages and can proactively see what you're talking about, see what you're struggling with, seeing what your schedule is, and helping keep you accountable to that schedule, and even seeing a missing task and be like, oh, hey, coworker.
Did you do this task? Oh, no. I haven't yet.
Claude can see that and be like, hey. Don't worry. I've done it for you.
So it's like literally having a coworker that you don't have to be in the loop for once you get to level three. You can see down here in plain English, it says it schedules and routines so that work starts itself. You need to turn on, of course, the security review automatically, the code review automatically, agent sandboxing so you're not overriding your actual product or whatever is live.
You, of course, need to set up a Claude MD file and skills to encode the standards. If you don't have a Claude MD file or skills set up yet, don't worry. It's very, very simple to do.
It's all literally plain text, and I have videos walking you through exactly how to do that. And once you have those guardrails set up, you are ready to ascend to level four of Boris Czerny's Claude Code users where only point 1% of people are.
Remember, Boris actually said that Anthropic as an organization isn't even in stage four yet, although Boris, as a high achiever of the class, is there himself. He calls it being AI native. I call it being an owner.
And in our pizza analogy, it looks like a thousand chefs across a 100 kitchens in the city. Most of them you've never met.
The kitchens hire their own chefs, so you're not even in the hiring process anymore. They restock their own fridges. They invent their own dishes.
You're not even writing menus anymore. They even open new locations while you sleep. Your job at god mode Claude code setup is you set the direction.
So you literally say, we wanna be the best pizza in the city, and the chain or all your AI agents, which this is where Boris gets into literally having thousands of agents work on your behalf, go and enact this for you. Your phone only rings or you are only needed not to approve work, but if something goes wrong, if there is a file or if something is not working in your business or whatever product it is that you're selling and getting ClaudeCodes help with.
What it looks like, the loop is fully closed and most agents are kicked off by Claude itself and not by you. Hundreds of thousands of agents run. You steer by intent and monitor by exception.
The quarter long migration becomes a workflow you kick off and check on. So literally something that would use to take you a quarter can now be kicked off by Claude and be running in the background and take zero work from you. The bottleneck now fascinatingly becomes identifying and automating work at scale.
So finding enough work and the right work to have this massive team of productivity to work on and enforcing the right guardrails for each type of work. So they're actually getting high quality work done. You don't just have a thousand chefs in there making absolute crap.
You want the world's best pizzas. Let's remember, this is our job. We are now set the direction.
Products that help. Chord Agent SDK to programmatically build and schedule agents because you're not even the one building the agents anymore. And once again, Chord Tag, active in most Slack channels, auto responding to posts, and of course, taking action on what it reads inside those Slack channels.
And again, Chordtag's weird because it's one of the most amazing features that has dropped. And again, I've done a video on Chordtag. But the problem with Chordtag is it's only based in Slack, and it's very limited.
It's only available to enterprise customers, and you can only use it and it can only act based on API credits, which gets really, really expensive, which is why I personally set mine up inside Telegram to do everything that Claude Tag can do and a lot more, but using Claude's actual subscription. If you wanna set that up for yourself and you wanna ascend to level three and level four with personalized help from me and my lead AI engineer, come and join the best AI community in the world.
The link is in the description. We'd love to see you in there. We have a massive discount for you going on at the moment.
So come and check it out and to send this letter with personalized step by step help where we can literally show you how to plug these into your software, how to set up the right loops for you and your personal situation. And the guardrails become, remember, at this point, because you have so much manpower and so much firepower, cost controls for automation so you're not spending crazy amounts.
But once again, we don't wanna be using API. This is why I highly highly recommend you just use your subscription so it doesn't cost you any more. And, of course, model selection for automation so that the right model is doing the right job so your bill doesn't explode.
But once again, if you just set it up like I've set it up and use your subscription, it's free. I always say free, but you're already paying subscription. So it's covered under your subscription.
Once again, I would love to help you ascend these levels. Click the link in the description below to come and join us inside of our amazing community. Thanks for watching.
I'll see you the
The Hook
The bait, then the rug-pull.
A tweet from Claude Code's own creator, viewed 1.4 million times, mapped out five stages of AI adoption — and by his own count, a third of users never get past the first one. This breakdown turns his table into a pizzeria you can actually picture.
Frameworks
Named ideas worth stealing.
03:05list
The Five Stages of Claude Code Adoption
Stage 0 — Gated (You: Cook): AI gives advice but never touches your work; bottleneck is permissions
Stage 1 — Assisted (You: Supervisor): one agent works but every change needs review; bottleneck is your attention
Stage 2 — Parallel (You: Head Chef): 5-10 agents run at once with auto mode on; bottleneck is reviewing before it goes live
Stage 3 — Supervised Autonomy (You: Manager of Head Chefs): dozens to hundreds of agents run continuously; bottleneck is trust in the loop
Stage 4 — AI-Native (You: Owner): thousands of agents run and start their own work; bottleneck is finding enough work to automate
A five-stage ladder for how much of the actual work an AI coding agent is trusted to do without a human in the loop, mapped originally by Boris Cherny and explained here with a pizzeria analogy — each stage swaps a specific bottleneck for a specific set of guardrails.
Steal forauditing how far your own team has actually pushed AI adoption, and deciding which single bottleneck to attack next
CTA Breakdown
How they asked for the click.
VERBAL ASK
15:14product
“come and join the best AI community in the world... personalized step by step help”
soft-pitches a paid community in the final minute after delivering the full framework breakdown, with the join link in the description
A breakdown of Claude Code's native /loop and /goal commands, shown live on a race-simulator agent and a newsletter-writing agent that grades its own drafts until they pass.
A YouTube host reverse-engineers Boris Cherny's one-line CLAUDE.md snippet, then runs a live three-way Claude Code test to prove a self-learning lessons file beats a static one.
A narrated screen recording, no code and no prompt, turned into a repeatable skill — demonstrated live on YouTube competitor research and a dinner order.
A nine-month-old Karpathy prediction about AI memory becomes Anthropic's new overnight consolidation feature — plus a free routine that replicates it for anyone not on an enterprise plan.
A creator runs the same workout-app build through Opus 4.8, Fable 5, and three advisor-mode hybrids to find the cheapest way to get Fable-level output without paying full price.
A tweet-reaction breakdown of Anthropic's viral five-archetype framework — and the Slack-embedded Claude agent quietly filling the sixth role nobody named yet.
00:12
00:24
00:43
00:51
01:24
01:26
01:43
01:55
02:03
02:17
02:27
02:44
02:53
03:04
03:18
03:35
03:45
03:58
04:22
04:38
04:42
05:00
05:13
05:26
05:42
05:56
06:12
06:20
06:34
06:44
06:56
07:18
07:34
07:49
08:01
08:19
08:21
08:42
08:45
09:02
09:15
09:44
09:51
09:59
10:10
10:31
10:36
10:47
11:04
11:17
11:24
11:37
11:52
12:09
12:18
12:29
12:48
13:11
13:16
13:30
13:43
13:55
14:05
14:24
14:30
14:49
14:56
15:06
15:30
15:39
15:54
16:10
12:27
12:53
Video of the Day15:19
09:00
09:41
10:42