Quick Tour of My AI Stack
Key Points
- The speaker walks through their personal AI workflow, highlighting each tool’s strengths, weaknesses, and workarounds in under ten minutes.
- They rely on **ChatGPT** (especially GPT‑5 “thinking mode”) for deep analysis and handling large context windows, but avoid it for drafting prose, PowerPoint, or high‑quality Excel work.
- **Claude Sonnet 4.5** is their go‑to for writing because it captures voice and follows detailed instructions, serving as a collaborative “thought partner” rather than a silent generator.
- The same Claude model also excels at Excel analysis, allowing the user to edit, explore, and generate data‑driven insights directly from spreadsheets.
- A newly released open‑source model, **Kimmy K2** from China, is presented as a promising PowerPoint‑generation tool, though the speaker notes current caveats and ongoing evaluation.
Sections
- Personal AI Tool Stack Overview - The speaker outlines their use of ChatGPT primarily for analysis and quick context digestion, notes its shortcomings for writing, PowerPoint, and Excel, and briefly mentions incorporating Claude Sonnet into their workflow.
- AI PowerPoint Creation Guidance - The speaker advises using Claude Sonnet 4.5 for secure, minimalist PowerPoint generation, warns that current AIs can’t achieve elaborate styles, dismisses benchmark‑driven tool choices like Kimmy K2, and suggests relying on ChatGPT for thinking and outline drafting when data isn’t sensitive.
- AI for Social Trend Mining - The speaker highlights a specialized AI (Grok) that excels at quickly gathering recent Reddit and X conversations about trending topics, while using a generative browser (Comet) for broader web tasks and automated LinkedIn messaging, noting its data‑in/out capabilities.
- Personal AI Development Stack Overview - The speaker extols Claude Code’s flexible, ecosystem‑rich capabilities while noting its action‑biased nature versus Cloud Code’s more thoughtful approach, urging users to pick the right tool for their AI workflow and sharing personal setup guides.
Full Transcript
# Quick Tour of My AI Stack **Source:** [https://www.youtube.com/watch?v=lY6voDZpu3Y](https://www.youtube.com/watch?v=lY6voDZpu3Y) **Duration:** 00:11:05 ## Summary - The speaker walks through their personal AI workflow, highlighting each tool’s strengths, weaknesses, and workarounds in under ten minutes. - They rely on **ChatGPT** (especially GPT‑5 “thinking mode”) for deep analysis and handling large context windows, but avoid it for drafting prose, PowerPoint, or high‑quality Excel work. - **Claude Sonnet 4.5** is their go‑to for writing because it captures voice and follows detailed instructions, serving as a collaborative “thought partner” rather than a silent generator. - The same Claude model also excels at Excel analysis, allowing the user to edit, explore, and generate data‑driven insights directly from spreadsheets. - A newly released open‑source model, **Kimmy K2** from China, is presented as a promising PowerPoint‑generation tool, though the speaker notes current caveats and ongoing evaluation. ## Sections - [00:00:00](https://www.youtube.com/watch?v=lY6voDZpu3Y&t=0s) **Personal AI Tool Stack Overview** - The speaker outlines their use of ChatGPT primarily for analysis and quick context digestion, notes its shortcomings for writing, PowerPoint, and Excel, and briefly mentions incorporating Claude Sonnet into their workflow. - [00:03:13](https://www.youtube.com/watch?v=lY6voDZpu3Y&t=193s) **AI PowerPoint Creation Guidance** - The speaker advises using Claude Sonnet 4.5 for secure, minimalist PowerPoint generation, warns that current AIs can’t achieve elaborate styles, dismisses benchmark‑driven tool choices like Kimmy K2, and suggests relying on ChatGPT for thinking and outline drafting when data isn’t sensitive. - [00:06:17](https://www.youtube.com/watch?v=lY6voDZpu3Y&t=377s) **AI for Social Trend Mining** - The speaker highlights a specialized AI (Grok) that excels at quickly gathering recent Reddit and X conversations about trending topics, while using a generative browser (Comet) for broader web tasks and automated LinkedIn messaging, noting its data‑in/out capabilities. - [00:10:10](https://www.youtube.com/watch?v=lY6voDZpu3Y&t=610s) **Personal AI Development Stack Overview** - The speaker extols Claude Code’s flexible, ecosystem‑rich capabilities while noting its action‑biased nature versus Cloud Code’s more thoughtful approach, urging users to pick the right tool for their AI workflow and sharing personal setup guides. ## Full Transcript
Today I'm going to do something I've
never done. I'm going to share with you
my personal AI tool stack all the way
through end to end what I use things
for, where they work, where they don't
work, where I'm frustrated, where I'm
also working around things so that you
get a sense of how this works. And we're
going to do it quickly. We're going to
do it in less than 10 minutes. So stay
with me. Number one, Chad GPT. Everyone
uses Chad GPT. What do I use it for? I
use it for analysis specifically. So, if
I'm needing something that has some
memory, that can handle a lot of
context, I don't run out of context
window, and I need it to think clearly,
not for writing, for thinking, for
getting the idea right, for being a
thought partner back and forth. Chad
GPT5 thinking mode is very, very useful.
What's interesting is that the auto
mode, the fast mode is becoming
increasingly useful for digesting large
amounts of context quickly. So, if I
need a rough pass, I can do that. But I
do not use it for writing. I find that
chat GPT can be used for writing if I
push it, but almost always I go
somewhere else because the default voice
for chat GPT is not good enough. I also
don't use chat GPT right now for
PowerPoint and I don't use it for Excel.
I don't find that the finished quality
of work is really there. I will use it
to produce CSV files if it's just a
simple one sheet spreadsheet and it's
got to be a table. Chad JPT does great
at that. Moving on. I use Claude Sonnet
4.5 a lot. I use it specifically for
writing because it's very very good at
picking up voice and at following my
instructions and actually getting voice
right. Especially if I give it a sample
with that voice. And so one of the
things that I like to do is I like to
ask it to brainstorm in my voice back
and forth with me. How can I write this
better? Can I tweak this paragraph? Can
I do a better job formulating this
point? It's a writing assistant. It
doesn't just sort of produce it for me
and I walk away. I keep emphasizing this
with writing. You are accountable for
every word you write. So, if you're
going to put something out there, you
better own it, however you made it,
whether you're writing with AI or
without AI. And so, I find it's a great
thought partner when I work with it that
way. Bonnet 4.5 is also what I go to
when I'm doing Excel analysis right now.
It's got a great tool set for that. I
have a whole guide out on that and I
find it super useful because I can
actually dig in and understand what's in
the box and produce a useful analysis or
edit an existing file. All of those
things are possible with Claude. That
being said, there's an even better tool
out there for PowerPoint and that just
came out recently and I'm really torn
about it. I'm going to tell you about it
and I'm going to give you the caveats
because this you're getting a tour of
the workshop, right? You get to see how
Nate does his personal stack. Kimmy K2
is an open- source model out of China
and they have done a phenomenal job of
putting together a PowerPoint skill that
enables you to make useful PowerPoint
presentations that look really good. And
I will show one with the Substack write
up. You can see it's a very simple
prompt, but you get a very useful
PowerPoint out of it. That being said,
because of the where the data is
located, you cannot really use this in
the US or the EU for corporate data.
It's just the protections aren't there.
And so if you're doing a fairly generic
presentation where you're not, you know,
you're using publicly available
information off the internet, it's
phenomenal. If you're using it for
personal use and you don't mind, it's
phenomenal. Great job. If you need
corporate data protections, I find that
Claude Sonnet 4.5 is super super useful
for PowerPoint creation. It's just a
little bit less designed. I tend to
prefer a spare, elegant, minimalist
approach with my PowerPoints. I wrote up
a whole guide on this I can link.
It's it's going to be what you need it
to be with with Claude. It's going to be
usable. It's going to be presentable.
But if you have a very elaborate
PowerPoint style, got to be honest with
you, you're not going to get what you
want out of Claude. You're not going to
get what you want out of any AI right
now. The AIs that are doing the work
with PowerPoint either have fairly
strong preset styling or they're very
clean and minimalist and elegant. One
thing I will call out is I do not use
Kimmy K2 for thinking. I don't use it
despite the benchmarks that it recently
took. like I find that it's just not as
useful in practice. This is why you
don't trust benchmarking and instead I
use chat GPT for thinking and I can pull
outlines and put them into Kimmy K2 if
the data is not sensitive. So that's the
PowerPoint piece. That's the Excel
piece. There's a little bit of the
writing piece in there. You may wonder
what about formatting word docs. Sonnet
4.5 is actually surprisingly useful for
formatting word documents. You just have
to ask it and kind of be specific about
what you want. And so that's another
useful tip. The weakness, and I know
people are going to call this out, so
I'll just say it. I struggle with it,
too. Claude has context window
limitations. In particular, if you are
building in a lengthy Excel or
PowerPoint skill, you will struggle at
some point with Claude running out of
context. How do I deal with that person?
Well, I wrote in my guide on PowerPoints
that that you want to think about
chunking the deck, right? You don't
necessarily want to generate the whole
deck in Claude at once. You want to
break it up into pieces, maybe five or
six or eight slides each. That's true.
You also want to separate out the data
and the narrative piece on bigger decks
so that you can get those right
beforehand and you don't burn tokens on
those in a debt creation conversation.
If I run into issues once in the chat, I
always go back and I restart the chat
with a smaller ask. I have very little
patience for running into issues more
than once. And so if I run into to
context window issues on claude, I'm
always going to condense the ask down
and go piece by piece and I will find
out where the context window runs out
and I will get usable information in the
meantime. And so yes, do I start again?
I absolutely do, but I'm very careful
not to repeat my mistake. I do not want
to go back and have to have multiple
times hitting the context window. So if
I ask for a PowerPoint and it hits a
wall at the end of the context window, I
never repeat that ask. If I do that with
Excel, I never repeat that ask. What
about search? So, with search, I love
perplexity. I find the perplexity is
really useful for most general purpose
searches. I love the research piece. Uh,
and I wrote about this recently, too. I
love the research piece. I love the labs
piece where you can dive in on discovery
and understanding and creating reports.
It's super fun, but it's not perfect.
And there is one use case where it
really struggles and that is finding
recent information on social networks
about a trending topic. It just not as
good. And so as funny as it sounds 4 is
really really good at just that piece.
And so I use Gro for that. I will go and
say tell me what people are saying on
Reddit on X about this particular issue.
Especially if you have like an AI topic
that is trending or a brand new uh
product launch and you want to know what
people are saying about it. Grock is
super super useful for digging in and
understanding that. I don't necessarily
trust it for larger scale thinking. I
don't trust it for outlining. I don't
trust it for general web research. But
for that particular thing, it is very
very good for finding social
conversation. What about web browsers?
So I am really torn on this one. My
generalpurpose web browser remains
Comet. So I'm using Comet a lot. I use
it and I like the data ins and outs it
has. I'm particularly fond of its
ability to do generative you. So if I'm
sending a message in LinkedIn, it will
literally generate a message pane for me
and let me approve a message I send and
all of that. That's great. And it has
data ins and outs to a few places uh
including LinkedIn where it's very easy
for me to not have to go. I got to be
honest, I don't love LinkedIn. I don't
like to be there. Having a data in and
out where I don't have to interact with
the site is fantastic for me. But beyond
that, it's very useful because it
combines perplexity search powers with
an agentic browser. And so it's easy for
me to get a chat next to the browser to
understand what the page is doing to ask
for the to do other things off page for
me, which I find really useful and to
keep the power of perplexity in the
meantime. And that's actually a use case
that I I don't know about you, but I
like the idea of the agent going off and
doing things for me. And comet really
sort of fulfills that. I think the Atlas
use case is really interesting. Atlas
came out recently. I have found things I
can do in Atlas that don't work as well
anywhere else, but that feels like I am
still figuring out where those use cases
are. I think one of the big differences
is this is a chat GPT first browser. So,
it brings my memories in. It is also
chat GPT first from a search
perspective. It is also chat GPT first
from an engineering perspective. And so,
you get all the strengths that that
comes with that the model remembers you.
The model talks to you like it knows
you. The model understands your
preferences. And in fact, the more you
use the model, the more it understands
how you think about the internet. So,
this is all good, but it also means that
you're going through chat GPT for search
versus Google. And that's a design
choice they've decided to make. It makes
sense from a product perspective, but
it's something you have to decide if you
want to live with. Um, and it also means
that you are sort of committed to their
vision of the agentic future, which is
frankly a little bit more button-down
and safety first. like they are keeping
the AI agent on the tab with you for
now. You were keeping an eye on it for
sensitive tasks. They have walled off
certain tasks that you can't do things
around banking etc. that make sense to
me like I wouldn't want it to be able to
do that initially. So they're assuming
that the AI agent is not entirely
trustworthy yet and designing around
that which I appreciate. One last piece
to call out on Atlas, it is it has the
chat GPT brain for code and that makes
it super useful if you are trying to
understand how to build something or how
to review something. So I used Atlas to
look at a GitHub repo and that was super
useful because I could pull out a lot of
the details of the code and get it into
a format I could understand and process
really quickly. I also used it when I
was looking at Lovable. So, if you want
to drive a build off of lovable using
Atlas autonomously, you can do that,
which is really cool. Now, we're going
to come to the terminal and the command
line. Yeah, you thought I was done. I'm
not done yet. I love both Claude Code
and Codeex. I find that Codeex is an
extraordinary strategic thinker in the
command line. I go to Codeex almost
daily when I'm trying to think through a
problem and I need succinct, clear,
strategic analysis. Codeex is also
really good at finding and fixing bugs.
I love that. that I can throw a really
messy repo at it and it just kind of
digs in and finds them. I love Claude
code because it is so sort of rich as an
ecosystem. I can tie in cloud skills. I
can tie in my local files. I love that I
can tie in the MCP servers that I want
to and I love that it's for lack of a
better term it has a friendly feel. Like
I love that it goes and does tasks and
checks back in. That being said, Cloud
Code has a very strong bias for action
and Codeex is more thoughtful before
engaging. And so you have to know which
one you're going to choose because
otherwise Claude Code is going to be
very tempted to just run. So that's my
personal stack. I got it to you in just
over 10 minutes. I would be curious what
your personal stack is. And of course
I'm going to get you uh all of the
guides that I've written that go with
this and also some extra details just
for this based on my own experience in
the last uh week or two because this
always changes and so I want to make
sure that you get the latest
information. Uh you're best informed.
Best of luck with your AI stack.