Latest / AI Fire Daily / #95 Robin: The Claude OS Blueprint - Turn Obsidian into a Voice-Controlled AI Command Center

Transcript

#95 Robin: The Claude OS Blueprint - Turn Obsidian into a Voice-Controlled AI Command Center · AI Fire Daily

0:00
We treat our digital notes like digital graveyards.
0:02
We just dump text in there and we never really
0:06
look back. But what if that wasn't the case?
0:09
What if your folder of simple text files wasn't
0:12
just a place where ideas die? What if it was
0:15
a living voice controlled command center? Intro
0:19
music. Welcome back to the deep dive. I am really
0:26
looking forward to this one. Today, we're exploring
0:29
a deeply fascinating concept. We are learning
0:32
how to turn the Obsidian app into a fully functioning
0:36
Claude OS. Yeah, I've been incredibly excited
0:38
to unpack this. It honestly, well, it completely
0:42
rethinks how we interact with our computers.
0:44
On a fundamental level, it moves us away from
0:47
just typing on keyboards all day long. You see
0:49
it all the time online. People build these incredibly...
0:52
pretty digital dashboards, they spend a whole
0:54
weekend making them look visually perfect. And
0:57
then they stop using them entirely after two
0:59
days. Right, because it just becomes more digital
1:01
clutter. Exactly. It looks like a cool spaceship
1:03
control panel, sure, but it doesn't actually
1:05
help you get your daily work done. Yeah, it's
1:07
just aesthetic theater. Aesthetic theater, I
1:09
like that. Yeah. This deep dive is about something
1:12
else entirely. It's about building a system that
1:15
actually saves you real time. It's about unifying
1:19
all your context into one single place. Your
1:22
files, your calendar, and your research all live
1:25
together in harmony. It keeps everything connected
1:27
in one central local hub. You finally stop jumping
1:31
between 10 different apps all day. Right. That
1:33
context switching is what absolutely kills human
1:36
productivity. So we're going to map out this
1:38
entire structure today, step by step. We'll explore
1:42
the foundational obsidian second brain framework
1:44
first. Yep. Then we'll decode the local voice
1:47
routing pipeline. Which is amazing. And finally
1:49
we'll build the automated skills. Yeah. Because
1:51
those automated skills are what do the actual
1:54
heavy lifting for you. It's a wildly powerful
1:57
sequence once you finally connect all those pieces,
2:00
but you really have to build it in the right
2:01
order. Yeah. You cannot command an AI if it has
2:05
no map. Right. It needs to know exactly where
2:07
things are located on your hard drive. So we
2:09
have to start with that foundational map. Exactly.
2:11
I think a lot of people misunderstand this part.
2:14
Obsidian doesn't give the AI magical human -like
2:17
memory. It just gives it a highly structured,
2:21
predictable map to follow. So let's talk about
2:23
setting up that specific map. The guide we're
2:26
looking at suggests a very clean three -folder
2:29
structure. Yeah. It's heavily inspired by Andrej
2:31
Karpathy's workflow. You basically rely on three
2:34
main folders for absolutely everything you do.
2:37
Just three. Just three. strips away all the unnecessary
2:40
complexity. OK, so the first folder is simply
2:43
called raw. I imagine this is just the unfiltered
2:47
dumping ground. Spot on. That's where all your
2:49
source data lives. Saved articles, unedited transcripts,
2:53
messy meeting notes. They go straight in there.
2:55
So no formatting. None at all. You don't organize
2:57
them. You just store them. Wow. OK. The second
3:00
folder is called the wiki. This sounds a bit
3:02
more structured. Yeah, this one is for your clean
3:05
synthesized reference notes. It's knowledge that
3:08
you have actually processed and refined yourself.
3:10
It's the information you might actually need
3:12
to reference next week. Got it. Then the third
3:14
folder is just called output. Right. That holds
3:17
your final polished deliverables. Finished reports,
3:21
final scripts, completed project files. Oh, I
3:23
see. They live exclusively in there. It's the
3:25
end of the assembly line. Organizing this is
3:27
kind of like stacking Lego blocks of data. Yeah.
3:30
You need a completely flat, solid base plate
3:33
first. Otherwise... Everything you build on top
3:36
eventually just falls over. That's actually a
3:38
great analogy. The structure keeps the data totally
3:41
modular and totally predictable. The AI always
3:45
knows exactly where the raw materials are. But
3:48
folders are definitely not enough on their own.
3:50
No. The AI needs a literal table of contents
3:52
to navigate them. Right. That is where the index
3:55
.md file comes in. Every single major folder
3:59
should have its own dedicated index file. So
4:01
it acts like a master directory for the AI. Exactly.
4:04
Both you and the AI know exactly what lives inside
4:07
that folder. Because if you have 3 ,000 text
4:09
files, this is utterly crucial. It is. It's the
4:13
difference between a clean system and a chaotic
4:16
data dump. The AI reads the index, and it instantly
4:19
understands the neighborhood. You also absolutely
4:22
need a file called cliude .md. Yes. This file
4:27
basically holds the core navigation rules for
4:30
the AI. It essentially tells the AI, here is
4:33
the sandbox, do not leave it. Right. It sets
4:36
firm boundaries and explains your specific personal
4:39
folder logic. You know, there's a real danger
4:42
here if you don't set boundaries. Yeah. I still
4:44
wrestle with prompt drift myself. Yeah. If I
4:46
leave the instructions even slightly vague, the
4:49
AI just kind of wanders off. Oh, totally. It
4:51
starts hallucinating entire directory structures
4:53
I didn't ask for. Right, and that's not just
4:55
annoying, it's actually wasting your processing
4:57
tokens. Exactly. Clear instructions in that rule
5:00
file stop the AI from guessing. It anchors the
5:03
AI to your actual existing reality. Once that
5:06
map is built and secured, we add the visual layer.
5:09
This is a custom Obsidian plugin that acts as
5:11
your daily dashboard. What's cool is the AI actually
5:14
builds this custom plugin for you. Yeah, you
5:16
don't have to code it yourself from scratch.
5:18
you decide exactly what shows up on that main
5:21
home screen. So it can show your API token usage
5:23
and your social stats? Yep. It can display your
5:26
calendar tasks or your morning news headlines?
5:28
You can even add direct buttons to run specific
5:32
automated skills. You get full terminal access
5:35
right inside of the Obsidian interface. That's
5:37
incredible. The visual layer helps you see what
5:40
is actually happening. You don't have to dig
5:42
through nested files or weird terminal outputs.
5:45
Right. It brings all that hidden automated backend
5:47
work directly to the surface. It makes the invisible
5:51
machine visible to you, then it's beat. Let me
5:53
ask you a question about this rigid, specific
5:56
setup. I know a lot of creative people resist
5:59
this kind of rigid organization. Sure. Does this
6:03
strict folder structure ultimately stifle human
6:06
creativity? I hear that a lot, but honestly,
6:09
no. Structure actually enables true creativity.
6:12
When you don't have to think about managing files,
6:15
your brain is totally free to just create. The
6:18
system handles the tedious management for you
6:20
completely. Rules actually give you more creative
6:23
freedom. Yeah, exactly. So now we have a wonderfully
6:26
organized map. We also have a clear visual dashboard
6:29
to look at. Right. If I still have to type out
6:34
complex commands, I haven't really saved time.
6:36
No, you haven't. How do we interact with it naturally?
6:39
We give it a voice. We stop typing every single
6:42
little command. Voice mode is where this system
6:45
starts to feel incredibly futuristic. You can
6:48
just talk to your computer organically like a
6:50
colleague. So you can ask it to pull up the terminal
6:53
while you drink coffee. Yeah. You can ask for
6:55
the biggest, most urgent thing on your schedule
6:58
today. Or you can just tell it to run the morning
7:00
trend report. It happens effortlessly. Let's
7:03
break down how this voice pipeline actually functions
7:06
under the hood. OK. It seems like magic, but
7:09
it's actually a very logical sequence. The first
7:11
step is obviously your spoken voice command.
7:13
Right. You speak into your microphone in your
7:15
normal everyday voice. Then a program called
7:18
Faster Whisper immediately takes over the audio
7:21
stream. I want to pause on that. What exactly
7:24
does Faster Whisper do in this specific pipeline?
7:27
It transcribes your spoken audio into text entirely
7:30
on your local machine. So it doesn't send your
7:33
voice to the cloud? No, it's completely private.
7:35
Okay. Then it hands that transcribed text off
7:38
to Haiku 4 .5. Right. This is a smaller, highly
7:42
efficient AI model acting as a traffic router.
7:45
Its job is not to be the smartest model available
7:48
in the world. It just needs to be incredibly
7:49
cheap and unbelievably fast. So it looks at your
7:52
text and accurately routes the request. Exactly.
7:55
Then, Cloud Code takes over to do the actual
7:58
heavy lifting. Yeah, Cloud Code executes the
8:00
complex skill or finds the right research report.
8:03
It does the deep thinking that Haiku isn't designed
8:06
to do. Finally, a program called Kokoro handles
8:09
the audio playback. Kokoro is basically a local
8:12
voice generation tool. Yeah, it turns the text
8:15
answer back into natural sounding spoken speech.
8:18
The entire goal of this specific pipeline is
8:20
sheer speed. Simple voice commands should feel
8:23
almost completely instant to the user. And to
8:26
manage that speed, the router system relies on
8:28
three distinct routing tiers. Tier one is designed
8:31
for running your pre -named skills. Right. That
8:34
tier is very fast. You trigger a save skill,
8:37
and the system reports back when it's totally
8:39
finished. Then tier two is for reading an already
8:42
existing metrics report. Yes. This is actually
8:44
the fastest tier of all. Really? Yeah. The data
8:48
already exists in your output folder, so the
8:50
AI just reads it aloud. Oh, I see. It doesn't
8:52
have to calculate anything new at all. That makes
8:55
sense. Then we have Tier 3. This is the catch
8:58
-all for complex, much deeper tasks. Right. This
9:02
is where you would start a headless session.
9:04
A headless session is an AI task running quietly
9:07
without a screen. Yeah, Tier 3 is intentionally
9:09
the slowest option. The AI has to do actual time
9:13
-consuming deep research or complex planning?
9:16
This brilliantly explains why some AI answers
9:18
are fast and others take time. Exactly. A report
9:21
question is fast because the data is already
9:23
sitting there ready. Right. Deep research takes
9:26
serious time because it literally starts from
9:28
scratch every time. The router handles all of
9:31
this complex traffic seamlessly behind the scenes.
9:34
Whoa! Imagine scaling to a billion queries. The
9:40
local processing power required is just staggering
9:43
to think about. It's a massive leap forward.
9:46
That actually brings up the most fascinating
9:48
part of this entire setup. The full voice pipeline
9:50
can run entirely on your own local machine. Yeah.
9:53
You aren't relying on a laggy cloud speech tool.
9:57
You are controlling your own local system purely
9:59
by your voice, completely offline. What are the
10:02
specific hardware demands for running things
10:04
like Whisper and CoCRO locally? You just need
10:07
a decent consumer -grade graphics card. A modern
10:11
GPU, which is a computer chip that processes
10:14
graphics rapidly, handles it perfectly. These
10:16
models are highly optimized for standard desktop
10:19
hardware now. You just need a modern graphics
10:21
card. Yeah, pretty much. Two secs silence. Okay,
10:24
so a map and a voice are pretty useless on their
10:26
own. True. The system needs to actually do something
10:28
productive for you. This brings us to the actual
10:31
engine of the OS. We absolutely need to talk
10:34
about skills and automations. This is what makes
10:36
the whole system truly fundamentally powerful.
10:39
Because a nice dashboard is cool to look at,
10:43
but workflows... are vastly better. Oh absolutely.
10:46
The real system starts when your daily work becomes
10:49
repeatable. Let's clearly define what a skill
10:51
actually is in this context. What? A skill is
10:54
a perfectly saved way of doing a very specific
10:57
task. So it has crystal clear steps, rich context,
11:00
and a strictly defined output format. Exactly,
11:03
and an automation is just that same exact skill
11:05
running on its own set schedule. Right. If you
11:07
only use vague prompts, the AI is completely
11:10
inconsistent. You ask it 10 times and you get
11:13
10 wildly different workflows. You have to turn
11:15
your work into very, very clear instructions.
11:19
Saying research YouTube trends is a terrible
11:21
vague prompt. Oh, it's the worst. It gives the
11:23
AI way too much wiggle room. A vastly better
11:26
version gives incredibly specific parameters.
11:29
Like what? You tell it to check the last seven
11:31
days of outlier videos in your exact niche. OK,
11:34
that's better. You tell it to group repeated
11:37
topics into actionable content angles. Then you
11:41
tell it to save the final report inside a specific
11:44
research folder. That gives the AI a clear step
11:47
-by -step process and a definitive end state.
11:51
It knows exactly what success looks like. How
11:54
do you actually find good practical ideas for
11:56
these skills though? Well, the guide suggests
11:58
a powerful concept called the combo method. The
12:01
combo method? Yeah. It uses your own memory and
12:04
your actual computer logs. It blends intention
12:06
with reality. First, you use a stream of consciousness
12:09
approach. Right. You literally just talk to the
12:12
AI. about your daily work routines. You verbally
12:15
ask the AI which parts of your messy day should
12:17
become automated skills. It helps you brainstorm
12:21
based on your feelings. Then you pair that soft
12:23
brainstorming with deep hard log analysis. Yes.
12:26
You ask the AI to review your past usage logs.
12:29
It looks closely at your last 90 days of actual
12:32
computer activity. It finds the digital tasks
12:35
you actually repeat the most often. I push back
12:38
on this a bit. Isn't checking your own logs a
12:40
bit overkill for a personal system? It can feel
12:43
that way. It feels almost invasive to analyze
12:46
my own computer usage like that. Why can't I
12:48
just tell it what I want to automate? I totally
12:50
get that hesitation. It sounds intense. But human
12:53
memory is deeply, deeply flawed. We lie to ourselves
12:57
about how we work. That's fair. Your brain gives
13:00
you an idealized intention. But those raw computer
13:04
logs give you objective reality. We love to think
13:07
we spend hours doing deep, meaningful research.
13:11
Exactly. But the logs mercilessly show we spend
13:14
hours just resizing images and formatting notes.
13:17
Ouch. Right. Using both methods together gives
13:19
you the absolute best skill ideas. There is a
13:22
golden rule for turning these manual skills into
13:25
true automations. Keep every single workflow
13:28
as a manual skill first. That is so important.
13:31
You have to run it manually a few times to test
13:33
it. You carefully check the output and face any
13:35
broken steps along the way. You make the output
13:37
format clearer and remove any lingering vague
13:40
language. Yes. You only fully automate it after
13:43
the output becomes completely boringly reliable.
13:45
Once you run a skill manually every single weekday,
13:48
then you automate it. A morning trend scan can
13:52
automatically run at eight every morning. Competitor
13:54
tracking can silently run automatically every
13:57
Monday morning before you wake up. Exactly. The
14:00
strict rule keeps your entire system running
14:02
cleanly. You completely avoid automating broken,
14:06
messy workflows. You only automate the tested
14:09
skills that already work perfectly for you. Let's
14:11
briefly talk about building the UI for all of
14:14
this backend work. You actually design the dashboard
14:17
visually first before writing any underlying
14:19
code. Yeah, you can use claw design or even just
14:22
find visual inspiration on Pinterest. OK. You
14:25
describe exactly what you want the interface
14:27
to look like. So you might ask for a clean, modern
14:29
interface with soft shadows. Right. You pick
14:32
the best visual mockup and let the AI build the
14:35
actual code. You also use an incredibly helpful
14:37
plugin called Hot Reload. Oh, Hot Reload is fantastic.
14:41
This helps the AI instantly update the code much
14:43
faster without restarting everything. It saves
14:45
so much time. There is also an advanced, deeper
14:48
layer called Loop Engineering. Yeah, this makes
14:50
the entire system self -correcting and self -improving
14:53
over time. So each automation... checks its own
14:56
final output against a preset goal. Right. It
14:59
evaluates its own work. If a trend report is
15:01
too weak or missing critical sources, the system
15:04
notices. Yep. It automatically improves the instructions
15:08
for the very next run. It learns from its mistakes.
15:11
You definitely do not need this complex loop
15:13
on day one, though. Right. You start with simple,
15:15
reliable skills and slowly build from there.
15:18
Quick question on that feedback loop. How does
15:21
the AI actually visually see the screen to fix
15:24
UI bugs? It uses advanced computer vision to
15:27
take actual literal screenshots. It compares
15:30
the visual screen to the original design goal.
15:33
If something looks broken, it completely rewrites
15:35
the code to fix it. It takes screenshots to compare
15:37
and correct. Yeah. It creates a completely seamless
15:40
automated feedback loop. Wait, songs are two
15:43
seconds. Okay. Let's unpack everything we've
15:46
thoroughly covered today. A real Claude OS requires
15:49
four distinct parts working together in harmony.
15:52
First, you have the custom visual command center.
15:55
This is your personal dashboard. Second, you
15:58
have the local voice mode for fast commands and
16:01
quick answers. Third, you have the program skills
16:03
and scheduled automations for your repeatable
16:05
work. And fourth, you have the clean, highly
16:08
structured obsidian map of your files. Exactly.
16:11
But the most important philosophy here is deeply,
16:15
truly personal. You absolutely have to build
16:17
the system around your own daily work habits.
16:20
Right. You should never just copy someone else's
16:22
aesthetic dashboard. Start with just one single
16:25
highly reliable skill. Make it absolutely perfect
16:28
and then scale your entire system from there.
16:31
That is exactly how it becomes a real living
16:33
working system. Yeah. It stops being a pretty
16:36
setup you abandon after a frustrating week. It
16:39
really forces you to deeply examine how you actually
16:41
spend your time. It beautifully highlights the
16:43
hidden friction in your daily digital routines.
16:46
Which brings me to a final thought for us to
16:48
ponder. Okay. If our personal operating systems
16:51
eventually run entirely on local self -correcting
16:54
loops, if they learn to know our actual work
16:56
habits vastly better than we do, where does our
16:59
active thinking end and the OS begin? Beat? Try
17:03
picking just one repetitive digital chore today.
17:06
Write down the exact steps to manually turn it
17:08
into a repeatable skill. Altioro Music.