Jack Roberts builds a personal BI dashboard for his entire AI stack — six pillars, eight dreaming dimensions, and a $20-a-month rebrand of Anthropic''s Dreaming preview.
Posted
3 months ago
Duration
Format
Tutorial
hype
Views
49.4K
1.1K likes
57 · 43
Big Idea
The argument in one line.
A self-improving personal AI operating system that aggregates your models, memory stores, and skills into one dashboard with nightly automated recommendations can optimize your AI spending and efficiency in ways individual disconnected tools cannot.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You're an AI builder or consultant already using Claude across multiple projects and want visibility into token spend, skill gaps, and ROI across your entire stack.
A solopreneur or agency owner running 3+ AI tools simultaneously and looking to consolidate insights into one dashboard to identify cost-saving opportunities.
You're experimenting with Claude's Dreaming feature and want a framework to systematize how self-improvement recommendations get surfaced and acted on.
SKIP IF…
You're still in the early exploration phase with a single AI tool — this assumes you already have multiple subscriptions, memory systems, and integrations to unify.
You're looking for a no-code, plug-and-play solution — this is a technical walkthrough requiring localhost setup, API integration work, and ongoing maintenance.
You're skeptical of Claude's Dreaming feature or don't have access to Anthropic's preview — half the system's value depends on overnight recommendation generation you may not have.
TL;DR
The full version, fast.
Disconnected AI subscriptions, memory stores, and skills waste money and hide opportunities, so you need a single dashboard that visualizes the entire stack and improves itself. The Claude Code Operating System unifies six pillars� models, plans, memory, skills, knowledge systems, and connections� into one localhost view, then runs an overnight Dreaming engine across eight dimensions (conversations, cost, skills, memory, sessions, workflows, external opportunities, business context) to surface four high-leverage recommendations each morning. Onboard by pointing it at where your models, data, and memories live, set your hourly rate so it can calculate ROI, and let it flag duplicated work, outdated memories, overpriced model usage, and skills worth building. The visualization layer is where agentic AI becomes manageable.
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.
Plugs Hermes/OpenClaude agents into the same OS roof. Shows schedule, sessions/day, top skills.
18:00 – 20:30
13 · Anthropic Dreaming positioning
''Anthropic just shipped Dreaming research preview — we don''t wait, we build.'' Re-frames product as cross-model, not Claude-only.
20:30 – 23:00
14 · Visualization layer thesis + soft CTA
Big-picture: visualization is the next trend in AI. ''How do I SEE the thing?'' Pitches skills video next.
23:00 – 24:45
15 · Wrap
Closes with a redirect to his next video on building powerful skills.
Atomic Insights
Lines worth screenshotting.
A personal AI operating system aggregates model subscriptions, memory stores, skills, usage data, and integrations into one visual dashboard — replacing disconnected pockets of intelligence.
The six pillars of a Claude Code OS are: models in use, subscription plans, memory architecture, active skills, knowledge systems, and integration connections.
A Dreaming feature that runs nightly on all your conversation history and usage data surfaces improvement recommendations you could never identify by reviewing individual sessions.
Knowing exactly how much each skill saves in time and money — and multiplying that by your hourly rate — converts an AI subscription from a cost into a quantified ROI.
Token limit visibility prevents the surprise of hitting a wall mid-project; knowing what's left at any moment changes how you structure and sequence sessions.
Surfacing outdated memory entries that are no longer referenced prevents the AI from acting on stale assumptions that have diverged from your actual current state.
A client-facing version of the OS dashboard that shows their ROI from deployed skills is a retention and expansion tool — it makes the value visible rather than assumed.
Self-improving systems that detect skill gaps from conversation patterns and recommend new skills to build are a qualitative step beyond static skill libraries.
Building your OS at the $20/month plan is possible but limited; the advanced multi-model usage and dream analytics require higher tiers where usage limits don't constrain exploration.
Anthropic's Dreaming preview is the conceptual foundation that Roberts extended into a full nightly self-improvement loop running on personal data rather than training data.
Opportunity gaps — patterns in your AI conversations that suggest a missing skill or unaddressed workflow — are only visible when someone aggregates data across hundreds of sessions.
An AI OS built on localhost rather than a SaaS dashboard ensures the intelligence stays local, private, and independent from any vendor's pricing or deprecation decisions.
Takeaway
Steal the ''personal AI OS'' frame.
Own-your-stack playbook
Every creator is paying for 5+ AI tools with zero idea what they spend, what overlaps, and what''s stale — there''s a category here, and the $6 Stack version of it is yours to take.
Pillar your product. Six named pains beats one generic value prop — Jack''s ''Stop overpaying for Opus / Always know what''s left / Catch stale memory / Kill skills that earn nothing / Retrieval that doesn''t embarrass / Reclaim dead integrations'' is a master class in turning abstract software into six concrete buy-buttons.
Frame it against the incumbent. ''Anthropic just shipped Dreaming — we don''t wait, we build it cross-model'' is exactly Joe''s ''rent vs. own'' positioning. Anthropic and OpenAI will ship narrow versions; the gap is the cross-stack, self-hosted, own-your-data version.
Make the dashboard the demo. The whole video is a tour of the working product — no slides, no whiteboards. The dashboard IS the marketing.
Build an ROI tile that lies a little. ''$241,426 net ROI'' is a fantasy number, but it sells. JoeFlow''s ''this week you dictated X minutes, saved $Y vs typing'' tile is the same idea — bake it in.
Ship the v1 ugly. Jack''s repeated ''I just sat down with coffees and went crazy'' permission-slip-style language IS the brand. Joe''s LFB Line is built for exactly this energy — use it.
Make ''dreaming''-style overnight passes a feature, not a roadmap item. A nightly cron that emails 3 high-leverage recommendations is a 200-line script — but it''s the single most-clippable feature of the entire product.
Don''t copy the Skool gate. The opportunity to beat Jack is to ship the actual code free / lifetime — let Anthropic + Jack rent the dashboard, sell yours for $49 once.
Glossary
Terms worth knowing.
agentic OS
A personal dashboard that aggregates an individual's AI subscriptions, memory systems, skills, and usage data into a single visual interface — inspired by the concept of an operating system but for managing AI tool stacks rather than hardware.
RAG (Retrieval-Augmented Generation)
A technique in which an AI model searches an external knowledge base for relevant information before generating a response, allowing it to draw on documents or databases not in its original training data.
Obsidian
A local, file-based note-taking application that stores notes as plain Markdown files on the user's computer — commonly used as a personal knowledge base that AI agents can read and write to.
Pinecone
A cloud-hosted vector database used to store and retrieve AI-generated embeddings, enabling semantic search over large collections of text — often used as a long-term memory store for AI agents.
Dreaming (Anthropic feature)
An Anthropic research-preview feature for managed agents that runs overnight memory and analysis passes, surfacing recommendations based on accumulated conversation and usage history.
token limit
The maximum number of tokens an AI model can process in a single context window, or the usage ceiling set by a subscription plan — hitting this limit can interrupt or prevent completions.
cache hit rate
The percentage of AI model calls where previously computed results are reused from a cache instead of re-processed, reducing cost and latency — a low rate suggests opportunities to restructure prompts for better caching.
vectorize
To convert text or other data into a numerical vector (an array of numbers) that captures its semantic meaning, enabling similarity-based search and retrieval in a vector database.
OpenRouter
An API aggregation service that provides access to multiple AI models from different providers through a single unified API endpoint and pricing interface.
visualization layer
A user-facing interface layer built on top of AI outputs or data stores that renders information visually — dashboards, diagrams, and interactive HTML pages — rather than returning raw text or JSON.
“As these models get better at creating things, the question becomes: how do I see the thing? What does it physically look like?”
thesis-level prediction about visualization layer of AI→ 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.
17px
metaphoranalogy
Claude Code AgenTic systems are the future and unlock capabilities that 99% of people don't even know exist. And in this video, I'll show exactly how to build a visual intelligence system for Claude.
This shows you every insight for skills, memory, cost, and how much money it's making you and your clients. And this intelligence system even uses data to improve itself and help you save more time and make more money, literally getting you light years ahead of your competitors.
And if you're new, my name's Jack. I built and saw my last tech starter with a gazillion customers. Now I'm building my own AI businesses and I just show you the stuff that actually works.
So if you haven't already, grab that beautiful coffee. Let's dive straight up. So this is the ClaudeCode operating system.
I'm gonna tell you exactly why you need to have one and how you build one that looks something a little bit like this and how it all functions together. Because basically, it carries all the information data that you didn't even know that you physically needed. And I'm gonna go through exactly what all this is, how this all connects together, and why it's so important.
So let's think about this for a second. The Claude code operating system. What does it work and what do I mean when I say Claude code operating system?
So the general idea of the OS, the operating system, is that we connect your world visually in one single location instead of having separate disconnected pockets of intelligence that are not basically thriving together and bringing one unique overview.
What we do is bring that all together. So we have all the models that you're using, everything you're connected to, all of your individual memory systems, local, Obsidian Rag, You know, you think of it, Kapathi rag, pine cone, all this stuff.
All your skills, your usage, all this intelligence together in one simple place that gives you and unlocks things that you couldn't really unlock outside of that. So it's one place for everything to flow together across everything from Claw, Chakrity, Codex, Gemini, DeepSeek. It doesn't matter what you're in.
This brings everything together and actually is the future of how we use models. So you can't see your own AI operating system. So there's a lot of stuff that you guess at.
Right? Like, what did I spend this month in tokens? Where can I save?
How much money am I making with this stuff? What does it look like? When will I run out?
What skills are worth my time? Is the model using outdated information and memory? What is connected?
Am I on the right tier? What are the opportunity gaps based on the conversations you're having with models? So what's really cool about the self improving memory system is that like, if you're talking about stuff with an AI and you don't have a skill, this will actually recommend to you specifics that you should be doing based on your conversations.
It's like a meta skill, and it's kinda linked to Claude's dreaming feature that they're gonna be releasing. I'm gonna show you how I think this takes a a cool step further using those kinds of principles. And what will this unlock?
Because it's like intelligence about how you're using everything with AI, which is really freaking cool. And the reason for this is that every builder, myself, you, we all build slightly differently. Right?
And therefore, our intelligence should be slightly different. What will be great for Jack might be different than what's great for you and your friend. So realistically, we need to get a good view of how you're using the models, what models you're using, where is your data, what's your intelligence.
And based on your own unique situation, we can find specific recommendations to help you take it to the next level.
And what's really cool about this whole system by the way is when you onboard clients, you can get this specifically for them.
And you can see ROI, how much time they're saving based on their skills. Right?
The skills that they're using in their system, all that beautiful information when we understand how much that time is worth. So it's really, really cool to to build this up for clients too.
So when we think about this, realistically, are six pillars that make the operating system in my estimation. And effectively what they are is the models that you're using, the plans that you're on.
So for example, I use ChatGPT, but am I doing that with an API or am I just paying $20 a month? We've got memory.
So how we store memory? Have we got it locally? Are we storing it in Obsidian?
Are we storing it in Pinecam? The skills that you're using on a day to day basis, your knowledge systems, and again, how we connect it all together. So the idea here with these pillars is we wanna stop overpaying.
So how are you being more efficient? What's really cool here is we'll understand what models you're using for tasks and say, hey, and give you dynamic recommendations to say, look, you should move to this model, should move to that model. You're gonna know what's always left.
You're gonna be able to find outdated memories that aren't servicing you anymore, that you're not referencing. You can kill the schools that aren't doing anything for you. And basically, just like really improve and level up everything together.
And you're always gonna know what's actually left. So it really helps you understand and manage your token limit so that you're not running out all the time. Right?
It's like no nasty surprises. You can see this in one cool holistic dashboard. And by the way, if you do want this full course, I'm releasing it inside my community, this v one alongside the full core code.
Masterclass, it covers everything from power features down to memory system, apps, build everything, design systems. So I'll put a link down below if you wanna go and grab the the v one, but you'll also understand how to actually build and what kind of goes in the information for this. So effectively, what we wanna do is capture how much money you're actually spending on AI on a monthly basis.
And what we can do is on the skills, okay, is understand the skills that you have and it knows this because it knows where the skills live on your laptop and I'll show what that looks like in a second. What it can actually do is figure out how much you're actually saving in your time and money based on those skills and what they actually drive for you.
But I'll I'll show more on that in a second. You've got your net ROI, which is how much this is making you minus your cost for things. We can see the subscriptions that we've got like a CallPro Max, ChatGPT, OpenRooters, and you can even see step by step your current limits for all of the models that you're currently using.
Clawd, ChatGPT, OpenRooter, etcetera. Now what's really freaking cool and my favorite thing about this that I built on this is the dreaming feature. So you have access to so much data on your computer.
Right? Like your usage, conversations you're having with AI, but no one's really looking macro level at like all of your thousand chats and all of your usage and saying, well, what can we learn from this?
Like, how can we actually be better? And what this does every single day, whether you touch it with your hands like this, whether you're on the beach, sipping margaritas, this finds improvements for you automatically.
And this refreshes every single day. And it also gives you a suggested ROI calculator on that in terms of how much time you'll save. So for example, you may say here, hey, your video scripts memory is two and a half weeks behind your work.
Right? Cool. Okay.
That's an interesting one. I can mark that as done. Awesome.
Or I can go to the next one. You're paying Opus prices for jobs that Haiku can do. Interesting.
And then why you're suggesting this, a little bit of information, and then specifically, how to action them. It'll show you stuff like you did the same competitive research four times this week. Interesting.
So I'm duplicating work. Title and intro, all joined together make it well, see, you get the idea. It's basically showing you dynamic things every single day based on how you're using AI on your computer and whether I'm using it in cloud or the terminal.
This brings it all together, which is really freaking cool. Then we've got all of our skills that we can check out and use. And we have our knowledge system which is cool.
And this knowledge system is freaking epic. I've had it say so myself because what you can really do here, you I can see everything that I'm connected to. And if I open up my knowledge system real quick, you'll see when I come down here, I can actually disable certain features.
If I just wanna look at Obsidian, for example, I can see the stuff that I've got in here that I'm like playing around with. I can bring up my local memory system, and then obviously just bring in Pinecone to get a full overview of what this looks like. And if I wanna take a different view to understand what's happening, I can see.
And I get my Pinecone memory system. Basically, it's like a visual overview. I'll have every memory that I'm using, local Pinecone and Superbase is connected in in one really nice crisp view and how they're all related to each other.
And we also add in here as well, by the way, is like recent activity. So if I vectorize information, if I like which just fancy way of saying store new stuff, I actually get a breakdown here of what's happening with everything that I'm doing, which is great. So it's just good to get nice little bits of overview on that.
So you're probably thinking, Jack, that looks really freaking cool. How do I actually set this up and use it on my computer?
Well, let me show you exactly how that works. Now, obviously, if you're in the group, can just let you come and grab this from the classroom if you like, so you can rock and roll. But what I've included in this is a onboarding wizard, and this will really explain to you how you set this thing on.
So we look at ClawdOS. I'm gonna go ahead and stop the whole thing. So effectively, what well, what this one does that I built is it also detects all of the different models and systems running on your computer.
So if it finds OpenCLR, ClawdCo, Clawd Desktop, OpenRooter, whatever it is, it will automatically show you these and you can toggle them on and off. So can pretty pick the ones that you have, which is freaking fantastic.
So first thing you need to do is basically help it understand where do the models live on your computer and what are you actually using right now. So that's kind of the first part to understand like what is the model architecture like in your world.
And it should be able to do this, but you definitely wanna prompt it this way. Obviously, this onboarding wasn't just for you, but you get the idea. This one's really cool.
Where do we store that data? Well, this freaking knows because it actually has access to it. So the next question is, of the things I use, where's all of the stuff living, basically, which is crazy.
Once you've got that, where do you keep your memories? So this one knows it can basically it's detected that I've got my Obsidian Vault and my Pinecone index because it's pretty smart and knows what's going on. But you can add custom folders.
Basically, the the fourth step is memory, which is where does your world live? Like, is it in Obsidian? Is it in Pinecone?
Is it in a different is it SQL? Is it in Superbase? Where is it?
So we can actually grab those memories and start to add context. Once we've done that, this one's really cool for me. I think it's really important.
We can't know the ROI unless we have an idea of what your time is worth or the employee whose job you are enhancing is worth. You So can literally write down how much your time is worth like this, and you can see actually it'll tell you based on the skills you've used in the past x number of days or once you create it, how much skills you've used, what your actual time saving actually is.
When that's done, we bring down to the the dreaming system. So again, what this actually looks like is eight different eight different kind of services, which I had to look at. So the eight dimensions of dream intelligence is what you sleep, dreams read your stack, and services, what's worth your attention.
So the idea is AI should work for us, like, even when we're not thinking about it. So it is conversation analysis. So read your last seven days of user messages, embed some clusters by intent, flags any tasks you've done manually three more times so it can it can suggest turning it into a skill.
And this is something that I picked up with Hermes. Do you remember the Hermes agent was really cool. Right?
Because it basically had this amazing skill of if if it identified that you something should be a skill, it would turn it into a skill. So I just built it into this operating system. And and this, by way, is by no means perfect.
I just kinda sat down with coffees and I went crazy, to be honest with you. Because I I love building stuff like this. But that's the whole point.
Like, it should kind of have this macro level overview of what you're doing and how you can improve it. We got cost intelligence.
So inspecting every model call looks for high opus usage on simple work, low cache hit rates, oversight contacts, recommends where you can be saving when based on what you're doing.
Then we've skill performance that tracks each skill's last use date, like if you've got antiquated skills you're not using is an opportunity to improve that. Memory health, session hygiene, like when it's gonna be on tokens, workflow patterns, external opportunities.
So it's looking actively for new skills or new updates that might fit your world really well. Like, you know, if you're building good design systems, for example, this is a free resource that I that might be useful for you. And basically, all this does, if I just pull this up right now, you can see, just basically codifies what brilliant design looks like.
So if there are other design repos, I would love to know about them. So it's good for me if I'm doing a lot of design work with Claude or Chattypical Codex. So it can actually surface those things to me and tell me what I need.
And also it looks at business outcomes and tries to understand context and what you're focusing in on. Then we've got basically, you just set the frequency. So once you've kind of designed your own dreaming system, which is like, how do I want it to think?
What kind of recommendations do I want it to give me? You codify that sort of stuff. Then you have the morning and evening, so how frequently when do want it to run?
Websites during dream is good because I like to personally enrich with new things. So this is, look, Jackie, you're spending quite a lot of time on this particular problem.
Let me go and get some data for you. Let me go and find out what the experts are saying about this thing. I I personally want to know that information.
And then there's this design thing, is creating cool ass images every time. Of course, we want that. Of course.
I mean, that there's there's not gonna add any value, but I just think the experience is important. So it's really cool. And again, you can actually add in your OpenAI key here and there, so it can actually do the generation for you as and when you need to.
Then pretty much after that, you're all set. So this one here, if you're using this wizard, you basically just copy this and throw it in, but it's actually linked up anyway. So it will do that for you.
And then, essentially, once that's done, you go through and then you have the dashboard ready to basically look at everything live for you. And in addition to your subscriptions, you got your token slash API equivalent, which is if you had used this with API, this is what you would actually physically be paying.
And then what I thought was really interesting as well on this outside of the memory system is getting an overview of the systems I'm plugged into, my connected knowledge system. So, like, actually, what's everything I've got, like, what is powering this dashboard right now? And if there's something that shouldn't be there, we can just add that in basically.
Or my schedule tasks, which are awesome. It even gives me an overview of my sessions per day, like, much I'm actually using it, and then skills that I'm using most likely at the bottom. And then on the settings, you can go through the wizard.
And then because we've now got this rise of personal AI agents, what I've also done to simplify this is you can add in your Hema's agent or your OpenClaw system actually within this called operating system so everything actually lives within within one roof essentially. Just making it, like, way easier to use.
And so Anthropic did ship Dreaming, but basically, it's research preview. So it's available now in research preview. Developers can request access from the core platform.
But guys, we don't wait for anything. We go ahead and we don't wait for Anthropic to release anything. We just go ahead and start building it.
When it does release, it's cloud only for managed agents, overnight memory pass, and it kinda sees everything. So I think it's a fantastic idea they've got. And why not just use it not just for Claude, but like everything you're doing on every app?
Like, shouldn't matter whether you're using codecs or anti gravity or clawed over terminal. It's all needs to be connected into one and tell memory system intelligence. Right?
So we've got eight buckets, four cards, one night. And again, our dream engine, our our kinda dream OS that we've got building here takes conversations, cost skills, memory sessions, workflow, external information, and business contacts into our beautiful dream machine to give you four high leverage recommendations just like that.
It's that freaking simple. As you can see, it's got all the different bits that you can down use. And, obviously, if you're building your own, feel free to imagine anything you like, but I just personally think these AI are really sick.
And this just connect to the biggest trend that we're going to see, and you have to get your head around, which is gonna be the visualization layer of AI. As these models get better at creating things, the question becomes, how do I see the thing? What does it physically look like?
So these are GenTic operating system, these dashboards are only gonna be more important, especially if you need to help clients or yourself understand what is the value of what we're doing.
And because we're encroached so many subscriptions and so many apps these days, we wanna centralize that and bring that into one place. It's really important. Knowing how to build the incredible operating systems is one thing, but building these super powerful skills is something completely else.
And we can learn that by watching this video.
The Hook
The bait, then the rug-pull.
Jack opens with a maximalist promise — agentic systems unlock things 99% of people don''t know exist — then immediately stacks a second promise: a visual intelligence dashboard for every AI tool you use, that improves itself overnight. Whether he delivers on that promise is the whole video.
Frameworks
Named ideas worth stealing.
04:00list
The Six Pillars of a Personal AI OS
Models
Plans
Memory
Skills
Knowledge
Connections
The six dimensions that define your personal AI stack — Jack uses these as the spine of the entire dashboard and as the framing for six pain-cards: Stop overpaying for Opus / Always know what''s left / Catch stale memory / Kill skills that earn nothing / Retrieval that doesn''t embarrass / Reclaim dead integrations.
Steal forany ''own your stack'' positioning — turn abstract product into six concrete pains, one card each
15:00list
8 Dimensions of Dream Intelligence
Conversation analysis
Cost intelligence
Skill performance
Memory health
Session hygiene
Workflow patterns
External opportunity
Business outcomes
The eight dimensions Dream reads overnight to surface high-leverage recommendations. Inspired by Anthropic''s Dreaming research preview but cross-model and self-hosted.
Steal forany ''AI works while you sleep'' product mechanic — give it dimensions, give it a frequency, give it a ritual
09:00concept
ROI Calculator (time-saved × hourly rate)
Onboarding asks for hourly rate, then each skill''s last-used + estimated time saved gets multiplied into a net-ROI number on the dashboard.
Steal forthe JoeFlow ''we saved you X hours this week'' panel — it''s the same math, and it''s exactly what creators want to see
CTA Breakdown
How they asked for the click.
VERBAL ASK
04:40product
“If you do want this full course I''m releasing it inside my community — this v1 alongside the full Claude Code Masterclass.”
Soft mid-video CTA + closing redirect to next skills video. The actual code is paywalled behind his Skool community — link in description (bit.ly/4tyq4Uz). Pitch is woven into the value rather than blocking it, but the full build is not free.
A 14-minute walkthrough for wiring Andrej Karpathy's self-auditing LLM wiki into Hermes agent — so your AI can read your inbox, meetings, and expert research, not just you.
A five-level walkthrough of building one locked brand system in Claude Design, then generating a burger-brand website, a 3D scroll showstopper, and a matching animated video from that same system.
Jack Roberts open-sources the seven-step Claude Code skill that chains Fable 5, Higgsfield, and a GPT-5.6 juror into an assembly line for animated product websites — for about $10 a build.
A five-level blueprint for turning one frontier model subscription into a personal operating system that remembers everything, debates itself, and keeps working while you sleep.
00:00
00:01
00:04
00:12
00:13
00:14
00:17
00:26
00:28
00:29
00:29
00:30
00:30
00:30
00:34
00:35
00:38
00:48
01:18
01:28
01:38
01:48
02:18
02:28
02:38
02:48
02:58
03:08
03:18
03:28
03:38
03:48
04:08
04:28
04:38
04:58
05:08
05:18
05:28
05:48
05:58
06:08
06:18
06:28
06:38
06:48
06:58
07:08
07:28
07:35
07:43
07:47
08:07
08:17
08:27
08:37
08:57
09:17
09:27
09:47
09:57
10:17
14:29
17:55
13:12
18:59
21:14
15:02