Latest / AI DTC WTF / AI DTC WTF: Stop Starting From Scratch

Transcript

AI DTC WTF: Stop Starting From Scratch · AI DTC WTF

0:02
If you work in D2C and you use AI and
0:05
you're wondering what the
0:06
F is going on every week,
0:08
this is your podcast, the AI D2C WTF
0:11
podcast, your home for
0:13
tactical tips, strategies, and
0:14
ideas that you can implement right now in
0:17
your AI workflows to make
0:18
your brand or agency more money.
0:22
Well, welcome to the AI D2C WTF podcast.
0:26
And let me just say by
0:27
starting, this podcast was really
0:30
for helping you get tactical information
0:33
about how to use AI better
0:36
in your D2C e-commerce agency
0:39
or brand, right? I mean, that's basically
0:41
what we're trying to
0:42
do here, isn't it, Will?
0:43
Yeah. I mean, very simple. How to make
0:45
your stack, whether it's generating ads
0:48
or just your stack in
0:49
terms of your operations better through
0:52
AI. So one of the things that
0:54
we have is we're interviewing
0:55
someone for most of the episodes that we
0:57
have, and we have a bunch
0:58
of episodes we've already
0:59
recorded. So be on the lookout for those
1:00
as they come. This is the
1:01
first one, but we're going to
1:03
try to, as much as we can as well, talk
1:05
about what's new in AI at
1:07
the time that we've recorded
1:08
it. There's just changes by the hour,
1:10
right? In the recent weeks
1:12
that we're talking about.
1:14
I mean, certainly we could talk about the
1:16
new model on ChatGPT or
1:18
what are we calling it?
1:20
GPT2 or whatever it's called for image
1:22
generation. So what do you think about
1:24
that, Will? And then what
1:25
are some other things that have happened
1:26
recently? Yeah. I was sort of bearish on
1:30
GPT's image modeling
1:31
and just GPT in general for quite some
1:33
time, but I think it's sort
1:36
of twofold. One, the GPT2 image
1:38
model sort of blows Nano Banana, the
1:41
Gemini Google model out of the water in
1:44
almost every stat. I ran a
1:48
bit of a cookoff in 15 different ad types
1:51
and GPT2 won all of them.
1:55
The caveats are it's a bit more
1:56
expensive, but ultimately, you know, in
1:59
terms of product fidelity, in terms of
2:01
matching your fonts,
2:02
in terms of matching your colors, it's
2:05
pretty unstoppable. And
2:07
obviously we're going to continue
2:08
seeing these components changing, but I
2:11
would say bar none, this is
2:14
the model to be using if you're
2:15
going to be generating static ads. And
2:17
then subsequently, if you
2:19
want to be animating those
2:20
ads, the newest model from Alibaba, Happy
2:24
Horse, I think is what it's
2:25
called. You know, who comes up
2:27
with these names? God only knows. But it
2:31
is, it's really quite good. You know,
2:33
Seed Dance from TikTok
2:35
was sort of reign supreme for a while,
2:38
but now sort of the Happy Horse model is
2:39
great for taking static
2:41
ads and animating them. I think so that's
2:43
good to know. That's a
2:44
that's a change that we've seen.
2:46
What other new announcements have you
2:48
seen that have sort of hit you and been
2:50
like, oh, wow, that's
2:51
that's something that's really cool. I
2:52
mean, one that I saw that I
2:54
thought was really cool is sort
2:56
of the operating system. It seems that
2:58
Jacob has put out that a lot of people
3:02
are curious and using
3:03
and it allows you to spool up agents. I'm
3:05
not really sure what it does
3:06
100%. Maybe you can help me
3:07
understand that. That seems to be
3:08
something a lot of people are talking
3:09
about. Yeah, I'm honestly not
3:10
too, too familiar with Jacob's setup.
3:13
Yeah. So he put out this thing that's
3:15
it's called HQ and it
3:17
allows you to share knowledge, skills,
3:20
workflows and API access across the
3:21
organization. So everyone
3:22
works at the same foundation. Nice. OK. I
3:26
mean, this sounds super promising. I
3:28
would say something we
3:30
struggle with internally and I get on
3:32
client and consulting calls and all the
3:35
time is where should I be
3:36
putting on my markdown files or where
3:38
should I be putting all my agents? You
3:39
know, right now they
3:40
just sort of live locally on my computer.
3:42
How do I share them with my
3:43
team? This is definitely a
3:45
pretty massive solve. I can't say that I
3:46
played with it, but it sounds pretty
3:48
promising, I would say.
3:50
Awesome. Yeah, I think it's interesting.
3:52
I mean, I think a lot of that is like the
3:53
next phase of AI seems to be the
3:56
organization of where we put all this.
3:58
And so that clearly makes a lot
3:59
of sense, which is pretty cool. Outside
4:02
of that, Claude, our best friend lately,
4:04
anything that Claude
4:05
has has put out that you've been like
4:07
absolutely loving or
4:08
utilizing recently? Yeah, I would say
4:11
sort of twofold. Obviously, the Claude
4:12
design came out. You'll hit
4:14
your limits pretty quickly.
4:17
The limits very fast. Yeah, very fast,
4:19
very fast. But we built our new website
4:21
on Claude design. It
4:22
took me a weekend to do totally nailed
4:24
brand guidelines. I would
4:26
say for one off projects like a
4:27
website, it's great for iterative
4:29
processes. So you prefer to use Claude
4:32
code, especially with
4:34
CRO landing pages. Claude code is still
4:38
bar none the best. If you
4:39
combine Claude code for Cell and
4:41
GitHub, you can spin up landing pages
4:43
really rather easily. The other thing
4:46
that is, and I'm sure most
4:48
folks listening to this may or may not
4:49
know about, is on Canva, they dropped a
4:53
tool that allows you to
4:55
upload AI generated ads. And what it will
4:59
do is it will take each layer of the ad
5:02
you created and separate
5:03
them. So you can move elements around,
5:05
you can change your copy. Folks have had
5:08
issues for ages, you know,
5:10
editing AI images. And so like that has
5:12
been a massive unlock for us, you know,
5:14
for, you know, for allowing our
5:16
editors to, to clean those minor things.
5:19
Be more flexible with it. Yeah, nice.
5:21
Cool. Totally. And then I
5:22
think, you know, the most
5:23
exciting announcement probably unrelated
5:26
to marketing was sort of the
5:28
announcement of the JV with
5:31
Anthropic, you know, with Blackstone,
5:33
Goldman, et cetera, sort of saying that
5:36
it's going to be spinning up this
5:38
service oriented consulting service to
5:43
ultimately implement AI within enterprise
5:47
businesses. And I think
5:49
folks have been talking about this for a
5:50
while and saying that service
5:53
augmented with AI is going to
5:56
lead the pack in terms of company types.
6:00
More, we had SaaS for a while, but now
6:02
it's service plus AI.
6:04
And this almost sort of confirms that
6:07
hypothesis that there is a
6:09
lot of demand for this and both
6:11
Anthropic and OpenAI, you know, OpenAI is
6:14
about to announce it as well. Doing this
6:15
almost confirms that,
6:17
like, if you are able to effectively
6:19
better augment your agency or your brand
6:22
with custom tooling with AI,
6:25
you are going to be at the forefront of
6:29
this, of this bubble, if you'll, if you
6:31
want to call it, even call it that.
6:33
Yeah, it's an, it's an exciting time. I
6:34
agree that, you know, you going through
6:37
this as a person that's
6:38
learning and listening to this podcast
6:39
and soaking up new knowledge is, you're
6:42
at the leading edge. And so
6:43
good for you for, for continuing to
6:45
follow along and yeah, seeing the big
6:47
player validated is huge,
6:48
right? So the more that you can help
6:49
people navigate this, the
6:50
more that you can experiment
6:51
yourself and be able to augment the
6:54
service offering that you
6:55
have is also a huge with this.
6:56
So it's a, it's a fun times. Our first
6:59
guest that we have on the
7:01
podcast is Grant Holschick,
7:02
who is really interesting. I mean, we
7:04
talked about a ton of Claude stuff in
7:06
this episode. We talk about a ton
7:07
of the structural ways that he's building
7:09
things. Thomas is also in
7:11
this episode outside of a cafe
7:14
in Prague, drinking a glass of champagne,
7:17
which is just like, so
7:18
you can tell by his vibe,
7:19
the questions are quite good. So let us
7:22
know what you think, Andrew at
7:23
foxwelldigital.com. I'm always
7:25
here to receive anything from you and
7:28
these episodes are going to keep coming,
7:30
right? So if there's people
7:30
you want us to interview, there's people
7:32
you want us to talk to, we're happy to,
7:34
and we're really excited to,
7:34
to dive into it. We go into these
7:36
episodes with a lot of
7:37
curiosity and that's our goal.
7:38
So without further ado,
7:40
let's take it onto the interview.
7:46
Thanks for coming on the
7:47
show. We appreciate it.
7:48
Yeah. Thanks so much for having me. I
7:49
thought you were going to
7:50
try to say my last name and
7:51
I was like, let's see how this goes.
7:53
Nope. I thought about
7:54
it and then I didn't.
7:55
And then I didn't. So I'm just going to
7:57
leave it for now. How do
7:59
you pronounce your last name?
8:00
Let's just see. Hoshik.
8:02
Hoshik. Okay. I always
8:03
thought there was an L.
8:04
So this is why this is the problem that I
8:06
have because I've said
8:07
it wrong to other people.
8:09
Hoshik. So Grant Hoshik on the podcast
8:11
and we are glad to have you.
8:13
By the way, so just like a 30
8:16
second brief intro, what do you do with
8:20
AI now? Yeah. I'm an AI consultant,
8:23
education, adoption,
8:25
implementation. And what that means is
8:27
that I am actively processing
8:30
transcripts, processing client
8:32
needs, processing my team's communication
8:34
into something coherent.
8:36
Right. So that could be
8:37
making skills for other people, making
8:39
skills for myself, making
8:41
scheduled tasks. It runs the
8:43
absolute gamut.
8:44
Well, so just, you know, you see a lot of
8:47
implementations in
8:48
business. And my first question is,
8:49
what are some of the biggest unlocks
8:53
you've seen for folks in like the last 30
8:57
days that you're that,
8:58
you know, you feel like every time you
8:59
show the super person, they're like, oh,
9:00
that's an insane idea.
9:01
I think it's a lot less sexy than people
9:04
will hope to hear. But really breaking
9:07
down the fact of like,
9:09
what a skill file is, when you think
9:11
about repeatability and
9:13
leverage in an AI world,
9:15
it all starts with a simple markdown
9:17
file. So like, I've spent
9:19
90 whole minutes explaining
9:21
what a skill file is and how to make them
9:23
and how to think about them
9:24
to companies. That has been the
9:27
breakthrough moment. But the more fun one
9:30
is that I've connected
9:31
Fathom, my note taker to HubSpot,
9:33
and every two hours, it updates every
9:35
single record that of people I talked to.
9:37
And that that's like
9:38
my happy place. I'm so happy.
9:39
What might be helpful is I think folks
9:42
maybe get a little confused about the
9:44
difference between like
9:45
a skill versus a markdown file. You know,
9:48
what are these
9:49
different components? Like,
9:50
as I understand it, and I'm, you know,
9:52
playing dumb for a reason, they're just
9:53
like a block of text,
9:55
right? Like, what exactly are these
9:57
things? And why would you not just use
9:59
Claude Chat directly?
10:00
And what is the
10:01
purpose of even using them?
10:02
Okay, so a skill file for someone
10:06
listening is it's
10:07
essentially like an SOP, right? It's a
10:09
standard operating procedure, but for AI.
10:12
And so there's two main parts to it.
10:14
There's the description,
10:16
which is formatted in YAML, which you can
10:19
just ask Claude to be
10:19
like, write this in YAML.
10:21
And then the real meat and potatoes of it
10:23
is markdown. And
10:24
markdown is the coding language,
10:26
the formatting language that Notion runs
10:28
on, ClickUp runs on. Anytime
10:30
you do like a pound symbol,
10:32
and then the text gets bigger, that's
10:34
markdown. Okay. And so what Claude does,
10:37
it takes your process
10:39
or your SOP, the way that you do things,
10:42
and turns it into this
10:43
file that it's going to follow
10:45
step by step when you request that output
10:48
or that skill. Okay. And so
10:51
most people in their role,
10:53
they have a collection of tasks and
10:54
deliverables that they're responsible
10:56
for. And you can make a
10:57
skill file for each of those
10:59
deliverables. Skill files can be updated
11:01
in any chat. You can ask a
11:04
skill to run. It can make an Excel file,
11:07
it can make a document, it can make a
11:08
PowerPoint, an HTML file.
11:10
And you can say, okay, I didn't love how
11:13
you did this part, just
11:14
update that part of the skill.
11:16
And then next time it'll get better. And
11:18
so why do people care about this is that
11:19
when you think about
11:21
Claude and co-work and Claude code, I
11:23
talked about leverage and repeatability.
11:25
So if you have a skill
11:26
file that accurately and consistently
11:28
gets you to a high quality
11:30
output in a 10th of the time,
11:32
then you can use that programmatically
11:36
through a scheduled task or
11:37
just in your chat, and you will
11:39
get the same high quality outputs more
11:42
consistently using less time. And so the
11:45
skill file is kind of
11:46
the key that unlocks AI leverage in each
11:49
individual role and
11:50
ultimately in a couple.
11:51
And for me, Grant, talking to you in
11:53
Lisbon a couple of months or like a month
11:55
ago, one of my biggest
11:57
takeaways is that one thing is the skills
11:59
files and using it for
12:01
yourself to, you know, make yourself
12:03
more productive. But you talked about
12:06
sharing skills with the
12:08
team. And for me, this was like,
12:10
I haven't really thought about it that
12:11
way before, but I've now
12:13
started doing it and it's like a
12:16
superpower unlock. But could you talk a
12:19
little bit about how that works and what
12:21
that is and maybe some
12:22
really practical tips on how to do that?
12:24
Yeah, absolutely. So why
12:27
do we share skill files?
12:29
Well, for one, if I'm making, let's say I
12:32
make a lot of branded documents. So
12:34
branded documents could
12:35
mean SOWs, scopes of work. It could mean
12:38
a prompt that I want to share
12:40
with one of our clients. So I
12:42
make a nice document explaining what is
12:43
the prompt, how to install the prompt,
12:45
how to use it and what you
12:46
should get from it. I want those to look
12:48
nice. And so this branded
12:50
document skill is something that I
12:51
had been using personally for months. And
12:53
then I realized that my
12:54
engineers are also making
12:56
deliverables and things that memos,
12:59
briefs, etc, for our clients to read. And
13:01
so I was like, okay,
13:02
well, we need to get you guys the same
13:04
level of output that I get.
13:06
But I don't need to teach you
13:07
what our brand colors are, you don't need
13:09
to necessarily know what
13:10
our hex codes are, because
13:12
all that stuff is inside the skill file.
13:14
So just by downloading that
13:15
skill file to my computer,
13:16
even though I had Claude write it, I've
13:19
iterated on it probably five
13:20
to 10 times, I downloaded it.
13:22
And then I gave it to someone through
13:24
Slack, I can give it to
13:25
them over email and say, Hey,
13:27
go into Claude, go to the customize tab,
13:30
click on skills doing this all like I
13:32
know exactly where all
13:33
the things are. Go to skills, press add
13:35
new skill, upload skill, bag the file
13:38
into that little box that
13:39
appears. And it'll upload to your, your
13:42
Claude database warehouse,
13:45
whatever you want to call it.
13:46
Now, if you're on Claude teams, you can
13:47
also go into that
13:48
customize tab, click on skills,
13:50
and then share a skill with your team. If
13:53
you are an admin, or if
13:55
your admin has enabled the team
13:57
members to share those skills. And the
13:59
last thing is that if you are
14:00
an admin, and you have a team
14:02
account, you can also create org level
14:05
skills that you would then
14:06
upload to the organization skills
14:08
that exists in your organizational
14:10
settings. That was a lot of ticky tacky
14:12
stuff. But you asked for the
14:14
practical. So I'm giving
14:16
you all the all the clicks.
14:17
It's good. I mean, I think what the this
14:19
is exactly what we want
14:20
to get into and give people
14:21
the idea of how they do this. One, I have
14:24
a question about Claude
14:24
code. I don't think a lot
14:25
of people have done a lot with Claude
14:27
code. Or, you know, it's
14:29
like people have built skills
14:31
potentially, right? Or they're working on
14:32
markdown files. But it's
14:34
like, what are what are some of
14:35
the like, top couple things that you've
14:37
seen you've done, or you've
14:38
seen other people do in Claude
14:39
code that you feel like are massive
14:42
unlocks for organizations?
14:43
So I'll say two things on this one, I
14:45
have a cool Claude code use
14:47
case. So I'll answer that part
14:48
first, which is, I think the the ability
14:51
to, for product people, engineers,
14:54
marketers, the ability
14:55
for you to visualize in an app format,
14:59
what's in your brain, like,
15:00
you don't need to storyboard
15:02
anymore, you don't need to go and like,
15:03
sit on the phone with a designer or
15:05
product person for an hour
15:06
and say, No, the button should be like
15:08
here, and it should look like an arrow.
15:10
You don't need to do
15:11
that anymore. You can go to Claude code
15:12
and describe what you want. And it'll
15:15
write the app. And it's
15:16
very easy to now deploy those apps to a
15:18
platform like Vercel. And I say that,
15:20
even though I'm an AI
15:22
consultant, like I took three coding
15:24
classes through my university time, hated
15:26
all of them, barely passed,
15:28
I didn't even pass some of them, still
15:29
graduated point being me and
15:31
code are like oil and water,
15:32
we don't get along. But I understand how
15:35
the things connect. And so
15:36
I'm actually more proficient
15:38
in a tool like co work that has the same
15:40
schedule task features that
15:42
has the same skill features.
15:43
So for someone that never really got
15:45
code, co work still does a lot of good
15:48
for me. But when I want to
15:49
get a thought out, that's much more
15:52
interface driven, like I want to click
15:55
around, I want to see
15:56
the thing lovable became really popular.
15:58
But like, I canceled my
15:59
lovable, because I know that I could
16:01
accomplish the same thing in cloud code.
16:03
So for marketers,
16:04
salespeople, product people,
16:07
if you felt like code was too far away
16:09
from you to actually be
16:11
able to do anything useful,
16:12
I can tell you that if I was able to just
16:14
describe what I wanted, and it made
16:16
something I think that
16:17
was relatively aesthetic, let's say, you
16:21
can take a screenshot of an aesthetic
16:22
website that you like
16:23
and give that to cloud code, and it'll
16:24
make something look pretty,
16:26
right. So just knowing that
16:28
that's possible, you can describe it and
16:29
create an app. I think that
16:31
that is that is a use case
16:32
that anyone should know that they have in
16:34
their back. Will, I'm sure
16:35
you have other notes on cloud
16:37
code, too. Yeah, I mean, I think it's
16:38
also worth calling out that like,
16:40
creating a markdown file,
16:41
or creating a skill is as easy as going
16:43
into your cloud chat and saying, you
16:45
know, maybe you go into
16:47
a cloud chat where you just had a very
16:48
long conversation about a
16:49
client, and you were going back and
16:51
forth, you know, maybe cloud chat was
16:53
helping you brief out an asset. And
16:55
you're like, Oh, God, I just
16:56
put so much good information in this
16:58
chat. Like, what do I do now? The
17:00
solution is as easy as saying,
17:03
Hey, Claude, can you write a markdown
17:04
file for everything we just discussed,
17:07
and make it generic,
17:08
so I can use it for any client, right? So
17:11
like, maybe I was using a
17:12
cloud chat for going about
17:14
creating emails. And like, I thought
17:16
like, Oh, this is such a
17:17
great way to unlock, you know,
17:19
open rates or whatever it was. Now I want
17:21
to apply that to any client.
17:23
So I can create this generic
17:25
markdown file based on the conversation
17:26
we have, we had. So Claude is
17:28
always sort of coming in with
17:29
that prior knowledge, right? And if
17:31
you're too scared to use
17:32
Claude code, you know, one, get over
17:35
it, it's not that scary. But two, you
17:36
know, your alternative is
17:37
you can use cloud chat, you can
17:38
take all these markdown files, put them
17:40
in your projects in cloud
17:42
chat. And like that works just as
17:44
well. Because I know cloud code can be
17:46
prohibitively expensive, because it's not
17:48
offered on the pro plan
17:49
anymore, you have to be on max. So if you
17:51
want to just do everything
17:52
that Grant sort of outlining
17:53
in cloud chat, you know, that, that
17:56
works, that works really well, you know,
17:58
as well. So I guess,
18:00
Grant, you know, I guess my question for
18:02
you is, you know, you're
18:03
chatting with a lot of companies
18:04
that probably are going from zero to one,
18:06
like, what is the first
18:08
markdown file you tell them to
18:10
create for their company?
18:12
Branded docs, branded PowerPoint, about
18:14
me, and voice and tone, those and about
18:17
my company. So I'll break
18:21
up. I love decks, PowerPoints. It's like,
18:24
in business, for some
18:25
reason, we've been obsessed
18:26
with like, what the 16.9 aspect ratio,
18:29
like, God bless it. Everyone
18:30
loves 16.9. I don't know why.
18:33
But you want to have PDF. Love it.
18:35
Everybody does. One time I was in a
18:37
meeting with a massive ad agency,
18:40
I was working and I was a consultant with
18:42
the brand. And they had a
18:44
consultant, like a massive
18:45
agency that came in and did this, I think
18:48
it was 300 slide deck.
18:50
And it was an all day like
18:51
presentation thing. And I found out later
18:54
that they were charging
18:54
them $75,000 a month. And all
18:57
they were doing was just like putting
18:58
this strategy in this deck. I'll never
19:00
forget that. I haven't
19:01
forgotten it. Obviously, this was like
19:03
six years ago. I mean, 75 grand for a
19:07
deck. So there you go.
19:07
There's a whole nother business line.
19:09
Anyway, I'm going to
19:10
reevaluate what I'm doing for a
19:12
profession. Anyway, the like people
19:16
really love decks. So understanding how
19:18
to make a really good
19:19
branding skill, I think is really
19:21
helpful. I talked about branded
19:23
documents, I make more document
19:25
files and I do decks. But I got on a call
19:28
with someone literally
19:29
yesterday, who's in the sales
19:31
team. And she's like, our sales team is
19:33
consistently pulling in different
19:35
features that we need to sell
19:37
to different enterprises. But we need to
19:39
have them all look the
19:40
same while they sub out
19:41
different slides, like section one versus
19:44
section three. And I
19:45
was like, okay, well,
19:46
we need to see what your template looks
19:47
like. And she showed me that she had
19:49
uploaded the template
19:50
deck as a PowerPoint, or sorry, as a PDF,
19:53
instead of as a skill. And
19:55
so AI, when you give it a PDF,
19:57
it has to use like optical vision,
19:59
instead of understanding what's in the
20:01
code, because PowerPoint,
20:02
Excel, Doc, those are all code based
20:05
things, which is why Claude
20:06
can make them with high accuracy.
20:08
So if you give it a PDF, you're going to
20:10
lose a lot of quality between
20:11
what your what your reference
20:13
is, and what's actually going to come
20:14
out. So branded decks,
20:16
really popular. And then just
20:19
general branding. So where your hex
20:21
colors, your fonts, etc, good to have
20:23
just across the board,
20:25
separate to the deck. And then I talked
20:27
about a few personal ones.
20:28
So your voice and your tone,
20:30
how do you write? How do you actually
20:32
talk to build out that
20:34
skill or that markdown file,
20:35
highly encourage you to get a tool called
20:37
whisper flow. Because
20:38
when you're talking,
20:39
not typing, you actually maintain your
20:41
authenticity and your voice.
20:43
And then you have your about me
20:46
file. Great, just a you can, I've got a
20:49
prompt if you want to email
20:51
me, but like, you can go and
20:52
just say, Hey, Claude, interview me to
20:55
understand who I am, what I do, how I
20:57
like to work. And Claude
20:59
will interview you and you can respond.
21:00
And that'll create an about me file to
21:02
give context about who
21:03
Claude is working with. And the last one
21:05
is about my company. And this
21:07
is one that should be shared
21:08
across the org. So if you want to have
21:11
consistency, share your about
21:13
my company file. And that way,
21:14
everyone will have the same reference for
21:16
Claude. I think that's great
21:18
examples. So let's get super
21:21
practical into this. I now understand I
21:24
need to create these
21:25
different documents. I understand I
21:27
can chat to Claude and he might do
21:30
something for me, but could you like,
21:32
break it down super simple?
21:34
What's the process of creating these
21:35
documents in as simple way as possible?
21:38
So the about me, the voice and the my
21:42
company will be more straightforward,
21:43
because you're not using
21:45
it. Like your reference is your is what's
21:46
in your head, right? So if
21:49
you start just to jump in there,
21:50
Grant. So when it's in your head, then
21:52
you would recommend
21:53
whisper flow versus writing,
21:54
because then it's more you use more
21:57
words, it's more colorful,
21:58
it's more in depth, right?
22:00
Yep, exactly. So in those situations, you
22:03
can start a new chat or a
22:05
new task and co-work. And just
22:08
explain what you're trying to achieve.
22:09
Like I want to create a markdown file
22:11
that you Claude will
22:13
reference when I start a new chat or
22:17
task. And this is specifically
22:19
like, this is really much more
22:20
useful for co-work than it is for chat,
22:23
because the chat is limited in the
22:26
knowledge base of a project,
22:27
or it's limited to your global
22:28
instructions. Co-work can actually
22:30
reference files like Claude
22:32
Code. So this is when we talk about file
22:34
creation, we're really thinking, okay,
22:35
well, how can we get
22:36
leverage on co-work and Claude Code? So
22:39
stating your goal, I want an about me
22:42
file that you can reference
22:44
when you're talking to me, interview me,
22:46
right? I also really like to
22:48
use a tool that co-work will
22:51
often recommend, but I asked for it
22:52
specifically, which is called Ask User
22:54
Questions. And so what
22:56
this does is it just pops up a little
22:57
multiple choice select. And
22:59
I like that for those kind of
23:01
quick, okay, well, how long do you want
23:03
the file to be? Or in what circumstances
23:05
will I be referencing
23:06
this file? Those types of things, like
23:08
it'll suggest answers that
23:09
you can just select as multiple,
23:11
multiple or single select and move on.
23:13
But then you want to start getting into
23:14
the whisper flow and really
23:16
speaking what's on your mind. And you can
23:17
go for two minutes of just yapping. And
23:20
that's great context
23:21
for Claude to have. So don't be shy,
23:24
really let it all out. And then it'll
23:26
make sense of that unstructured
23:27
text. And so you can do that for your
23:29
voice, your about me, and your my
23:32
company. Those those three are going
23:34
to be really good for the self interview.
23:37
I've done, I think the my
23:39
about me, I spent literally
23:41
like an hour back and forth until it felt
23:43
good. I had a really good
23:44
starting prompt to work from
23:46
that was like probably too comprehensive.
23:48
But, you know, let it take 10 minutes,
23:51
let it take 15 minutes
23:52
or 30. Like it's a it's time well spent,
23:54
because that's really strong
23:55
leverage for when you're working
23:56
with Claude. So just to follow ups from
23:58
that. So one thing is you created the
24:00
file, you start using it.
24:01
But as I understand it, you can also kind
24:03
of ask it to update itself.
24:05
And and it's it should be like
24:08
a living document. Am I right? Yeah,
24:11
definitely. All I think all of your
24:12
files, all of your markdown
24:13
files can be considered living documents.
24:15
Because at the end of the
24:16
day, this is what I learned from
24:18
automation, like edge cases exist in
24:20
life, just generally, right? So if I'm
24:24
making an SOW, and we
24:25
have these core offers, and then someone
24:27
comes to me and says, all right, well,
24:29
I'll pay you triple your
24:29
rates if you do it completely
24:31
differently. And I'm like, Okay, well, I
24:32
want to start from this
24:33
template that I have. But let's work in
24:35
this new option, right? I might want to
24:38
use this new service
24:39
again. So you can then say, go to my SOW
24:42
skill and update it with this new this
24:44
new path, potentially.
24:46
And so even when you think about your
24:47
about company, like if
24:49
we another great example,
24:51
we're growing, we're hiring, new person
24:53
joins the team, part of the
24:54
about company tells you what the
24:56
team roster is and what each person's
24:57
role title is and what they're
24:58
responsible for. So if you hire
25:00
a new person, great idea to go and update
25:02
that file, simply by starting
25:04
a new chat thread and saying,
25:05
update my, my company file, because I've
25:08
just hired Bennett, Bennett runs media
25:11
buying, senior media
25:13
buyer, he's going to run these five
25:15
accounts. Great, that's gonna be super
25:17
helpful for Claude to know
25:18
that. That's great. And then my second
25:19
follow up would be, so
25:21
that's kind of the the markdown
25:23
documents. So what about like the designs
25:25
and getting it to make the
25:27
correct presentations and
25:28
all that stuff? How do you go about that?
25:30
Yeah, so skill files, for
25:32
the most part, are going to be
25:33
just single markdown files, single text
25:35
files. However, when we want
25:37
to have reference material,
25:39
like a PowerPoint template or an Excel
25:41
file as a template, we might
25:43
want to have our skill reference
25:46
that material. This is a great example
25:48
for our decks, right? And so
25:50
if we are in the process of
25:52
creating a new skill for branded decks,
25:55
then we can provide that
25:56
PowerPoint file to Claude.
25:58
Claude will read the file, understand
26:00
what's there, give it as a PPTX, not a
26:03
PDF. And then when we make
26:04
a skill file, we can tell Claude, include
26:07
this PowerPoint file as
26:09
reference material. And the
26:11
big difference here is that at the end of
26:13
its process of asking
26:14
questions and getting answers
26:16
and writing that skill file, there's
26:17
going to be multiple files that it spits
26:19
out. And you're going
26:20
to want to click download all because
26:22
then it's going to zip those files
26:24
together and you can move
26:26
the entire zip folder into Claude. And
26:29
when you upload that, all
26:31
the material is there. It's
26:32
referenceable and it's a larger body of
26:36
work for that skill to work off of. And
26:39
that's where you get
26:39
really specific and granular skill files.
26:44
Great for branding,
26:45
great for visual things.
26:47
One thing I've always wondered that I
26:49
haven't had the guts to
26:50
ask someone is why and do you
26:53
recommend setting up like a dedicated
26:56
computer for all of this? Like, I mean,
27:00
because I got a thing,
27:01
you know, dispatch or whatever it is on
27:02
Claude that was like, hey, you can do
27:04
this from your phone
27:05
and like, it's still running on another
27:06
computer. And I don't, you know,
27:08
apparently people just have
27:09
computers sitting around. Like, I don't
27:11
have that kind of money. Just
27:12
like having a computer sitting
27:13
around. But what do you recommend doing
27:17
that as you continue to sort
27:18
of build on this? I guess I'm
27:20
just curious of your opinion about the
27:21
infrastructure side of it. I'm getting
27:23
into that much more right
27:24
now. I kind of went down the rabbit hole
27:26
of trying to push co-work as far as I
27:28
could with this idea of
27:30
like a helpful employee companion. And so
27:32
what I did was for the
27:34
longest time, I've had scheduled
27:36
tasks working in co-work that read my
27:38
transcripts, Slack, email, look up and
27:42
identify open loops,
27:43
things that I said I would do, but I'm
27:45
not doing in the call. And it logs those.
27:47
And I found out that
27:48
I had six different areas in my computer
27:50
that that information was
27:51
being logged. So that's a no,
27:52
no, don't do that. But then I was like,
27:54
okay, well, we got to
27:55
start cleaning this stuff up.
27:56
And so I thought, okay, well, if I had
27:58
co-work go and index that list, try to
28:01
fully execute a task.
28:04
And if it couldn't give it up to someone
28:06
on my team and say, I
28:07
need information from you.
28:08
But an actual employee would first go try
28:10
to see if the task had already been done,
28:12
because maybe I do it on a call, but the
28:14
transcript doesn't say
28:14
that. Or maybe I sent the email,
28:16
but I never confirmed with Claude. So you
28:19
need to go look for
28:19
proof of completion. So
28:20
where I'm going with all this is that
28:22
co-work has limits.
28:23
Scheduled tasks only work on time.
28:26
They don't work on tool. They're not
28:28
tool-based triggers. So for example,
28:31
difference is when Fathom completes a
28:33
transcript, it is not going
28:34
to start a scheduled task. A
28:36
scheduled task can run every two hours
28:37
and look for calls that
28:39
finished in the last two hours,
28:40
but Fathom can't start the workflow.
28:42
Okay. So from an
28:44
infrastructure perspective,
28:45
I am now taking a playbook, a note out of
28:48
a play out of Thomas's
28:49
playbook. I've got a Mac mini
28:51
that I just turned on today. And I am
28:54
currently moving all of the
28:56
edge pushing, the boundary
28:59
pushing that I've done in co-work. I'm
29:00
moving that over to my Mac mini. I'm
29:02
setting up OpenClaw and I'm
29:04
setting up the Gary Tan infrastructure
29:06
for file storage. And so I'm combining
29:09
all these things now
29:10
into this external thing. Main reason why
29:12
my computer is on fire all day. And I
29:17
cycle through the battery.
29:18
It is a less than one year old MacBook
29:20
Pro. And I cycle through
29:22
the battery five times a day.
29:23
That cannot be good for me. Like I just
29:26
doesn't stand to reason
29:28
that that's a good idea.
29:29
So I moved it to a new computer. I
29:31
definitely, I definitely identify with
29:33
this. Uh, this is why
29:34
I asked because like I asked, it was
29:36
doing some really
29:37
complicated shit the other day.
29:38
And like at the same time was also
29:40
building a, um, monkey riding
29:42
a toilet cart game for Nora.
29:45
By the way, it's pretty dope game. If you
29:46
want to play, let me know.
29:47
You got a lot.
29:48
What's the context though?
29:50
Uh, she, I asked her to come up with a,
29:52
um, a funny game and she came
29:54
up like an idea. So I'm trying
29:55
to teach her like game design and design
29:57
thinking and stuff. And she designed
29:59
basically this Western
30:00
themed game where a monkey rides a toilet
30:02
around space. It's very fun. And it's,
30:05
uh, you have to visit
30:06
different like Western themed bathroom
30:09
stops. The monkey does. So
30:13
she's sick. That's amazing.
30:16
I'll send it to you guys. It's pretty
30:17
dope game. But anyway, so, so
30:19
we're like, I've got this and
30:21
it's like running this other, like two
30:22
other things that are like quite
30:23
complicated. And I mean, it was,
30:25
it took like the, all of it was taking
30:27
like 20, 30 minutes and it
30:29
was comical because my, it was,
30:30
my computer was flaming hot. I'm like, I
30:32
think I'm reaching this
30:33
point. And so I didn't like, I,
30:36
that's, that's why I asked
30:37
the question. But anyway, yeah.
30:38
Just a comment, Grant. So first of all,
30:40
welcome. And it's, it's
30:41
great to have you in the,
30:42
in the cloud community. Uh, what I think
30:45
going back to the start of
30:46
your point of like building,
30:48
building your skill bases and your MD
30:49
files, right? So what's
30:51
amazing with, um, AI going forward,
30:54
for example, with the open clause where
30:55
you can install it on your
30:57
Mac is that I use a shared
30:59
library via Dropbox. So everything that
31:02
co-work or cloud code or cloud chat is
31:06
learning and updating
31:07
these MD files of the company about me.
31:10
I'm also putting that into
31:12
my open clause agents who are
31:15
also reading and writing from it. So all
31:17
the infrastructure that
31:19
you have created with cloud,
31:20
you can just directly put into kind of,
31:23
uh, the
31:24
infrastructure can keep working in,
31:26
in another platform. And I think that's
31:27
the magical part about
31:29
MD files and skill files.
31:31
Yeah. I'm setting up that sync tonight,
31:33
late, late in the evening hours.
31:35
And then just to, and just to add onto
31:37
that, like, so you're new to this
31:39
ecosystem. One thing I'm
31:41
curious about you guys thoughts on is
31:43
just like co-work. Like I don't use
31:44
co-work at all. I just
31:45
use cloud code for everything. Is there
31:46
like any advantage at all in using
31:48
co-work, uh, versus using code?
31:51
Uh, that's my first question. Second is
31:53
just like, sort of like a
31:54
piece of advice for our listeners
31:56
here is like your first mark dial
31:58
markdown file or your first skill or
32:00
whatever the fuck you want
32:01
to call it will probably suck. Like it
32:03
may not be perfect. And so
32:04
what you need to be doing is
32:06
iterating. Right. And so like internally
32:09
we have marked on files for every
32:11
different type of asset.
32:12
So if I wanted to create like a before
32:14
and after ad and I say, Hey, use this
32:17
markdown file to create a
32:18
before and after ad for me, for this
32:20
brand. And it does, it generates a prompt
32:22
for me. And I need to
32:24
put that prompt into nano banana, GPT
32:26
two, whatever it is. And I get an output.
32:28
What I need to do is I
32:29
need to take that output and I need to
32:31
put it back into code. And I need to tell
32:33
code, these are all the
32:34
issues with this ad. You need to go back
32:37
or maybe more politely, please go back
32:39
and update all of the
32:40
respective markdown files that you use to
32:43
get to this point. So we
32:44
don't get this mistake again.
32:46
So the idea is like you're creating an
32:48
iterative loop. And I will
32:49
say, I may be, you know, I may be
32:52
on an Island in this group here. I will
32:54
say having a human in the loop doing that
32:56
is far more beneficial
32:57
than having a QC agent do this work. Like
33:00
I've tried having a QC agent
33:01
just like iterate and iterate
33:03
and iterate. It's fine, but it doesn't
33:05
necessarily get there. The
33:06
one thing that humans have as an
33:08
advantage, especially in the marketing
33:10
space is taste. Like, you
33:11
know, and we know on this call,
33:13
you know, what maybe like a good ad looks
33:15
like or what a good
33:15
landing page looks like,
33:17
right? Like we can optimize as much as
33:19
possible, but like there's, there's,
33:21
there is an innately
33:22
human intrinsic thing about taste. And I
33:25
think that's why, you know,
33:27
having a human in the loop
33:28
do this thing. But so anyway, that was
33:29
just a bit of a tangent and an aside. I
33:31
want to go back to the
33:32
cowork versus code thing, because maybe
33:34
I'm missing something here. And maybe our
33:35
listeners are missing
33:36
something here because I think there's a
33:38
code seems to solve any
33:39
problem that you could ever do.
33:41
Can I, can I go first? I have, I have
33:43
feelings about this. So I've been
33:45
experimenting with chat,
33:47
org, cloud, and open cloud. And it's,
33:49
it's a, it's a strange harmony and
33:51
there's a lot of feelings
33:52
around it. But for me, what I landed on
33:54
is that a cowork is my product manager
34:00
that connect. It's easier
34:02
to connect to all the different
34:03
connectors. It's easier than to just make
34:05
sure it has all the files
34:07
connected. Cloud code for me is too
34:10
complicated. It looks cool when my wife
34:12
looks behind my back on
34:13
what I'm doing, but in terms of just
34:15
making it do consistent,
34:18
good quality work for me,
34:20
cowork is better. Especially in terms of
34:23
making sure all the connectors and
34:25
everything is working
34:26
because on cloud code, I have a hard time
34:29
making sure it remembers
34:31
that it has to do this connector
34:33
or the liner connector or whatever. So
34:35
for me, it is my project
34:37
manager that I talk about when I'm
34:39
going to develop different things. And
34:41
then I get that to give me
34:43
the prompt I need to do in cloud
34:46
code to develop. So that's my take on it.
34:49
Got it. That makes sense. Maybe it's
34:51
gotten better since the
34:51
last time I used it, but I've just found
34:53
that using APIs and code
34:56
is just like so much more
34:57
efficient than the connectors and cowork,
34:59
because I feel like the
35:00
connectors, and again, it's been a
35:02
few weeks since I've done it last, but I
35:03
felt like the connectors just kept
35:04
breaking down. And then I
35:06
have to go reconnect them, reconnect
35:08
them, reconnect them. I was like, this is
35:09
not a repeatable system.
35:11
There's a very known bug about this
35:13
connectors dropping down with the
35:15
authentication thing
35:17
going, which I think will be resolved
35:19
soon. But you're correct.
35:21
Some of them are breaking down
35:22
and it is much better to connect directly
35:25
to the APIs. But I think for
35:27
the normals, the normies who
35:30
are just using this every day, I think
35:31
it's easier to just understand you have
35:33
cowork, you put in all the
35:35
connectors and it will just work for you.
35:37
And of course, they will prompt you if
35:38
you need to reconnect
35:39
or re-authenticate things. But it depends
35:42
on the nerdiness level, I
35:44
think, of the whole delivery.
35:46
But for me, showing cowork to different
35:49
clients, which is like small e-commerce
35:52
businesses in Norway,
35:53
they get their mind get blown by just
35:55
understanding how cowork works. And, you
35:59
know, applying all the
36:00
things Grant have been talking about
36:02
today and just giving them MD files and
36:04
creating schedules and
36:06
stuff like that. You are like the coolest
36:08
person in the office. You're like a
36:10
superhero. You're like the
36:11
the peak AI person just applying these
36:15
simple tactics, I think.
36:17
And I'll throw in on that completely
36:19
agree. The percentage of
36:20
people that are going to enjoy
36:22
code, cloud code is disproportionately
36:24
smaller than people that are going to
36:26
jump in and immediately
36:26
get value from cowork. So if you're
36:28
listening to this and you're like, I
36:30
don't know what to do here.
36:31
Go for cowork, have a ball. Right. And to
36:35
your point, Will, is this a
36:36
video podcast or is this audio?
36:37
We have video component as well on our
36:39
YouTube channel. Yeah.
36:40
Fantastic. Do you mind if I screen share
36:41
something? No, let's do it.
36:43
Okay. So this is what I was looking off
36:45
to the side and in eager anticipation.
36:50
So this is something that we teach a lot
36:52
in our workshops. Super
36:54
colorful. Effectively,
36:56
to get a good skill, I tell people it
36:58
takes three iterations. But
37:00
when we say iterate, it's like,
37:02
okay, well, what does that mean? Right?
37:03
Like really tactically,
37:05
brass tacks, how do I do that?
37:07
So what I tell people is take, you need
37:09
to have three inputs that are very
37:10
similar. So let's say
37:11
we're trying to make a Excel file of
37:14
financial report, right? Or
37:15
we're trying to do some sort
37:15
of manipulation. You can have input one
37:17
and you're going to prompt and you're
37:18
going to say, all right,
37:19
I really want this to actually be a
37:21
cleaned up Excel file with my brand
37:23
colors. And I want the
37:24
tabs to be really nice. The first tab
37:26
needs to say, start here. And then
37:28
everything else needs to be
37:29
logical from there. Give it your best
37:31
shot. Cloud's going to make an output.
37:32
You're going to say,
37:33
you missed the mark about 80% of the
37:35
time. And then when you
37:37
revise it, you say, okay, great.
37:39
I've revised what you worked on. Make me
37:41
a skill to do this, like my
37:42
revision next time. You take that
37:44
and you go to input two and you say,
37:47
great. Hey, here's input two. Use the
37:49
skill I just created.
37:50
That's going to make another cloud
37:51
output. You're going to revise that
37:53
because it's going to miss
37:54
something, but it's going to be much
37:55
closer, much better. And then
37:57
you get the two of that skill.
37:59
And then lastly, you take input three,
38:02
which is unique from the
38:03
first two, apply skill v2 to it.
38:05
You get an output. Maybe you need to
38:07
revise it, but after you
38:08
revise it and you say, great,
38:10
update the skill to my revision, that
38:12
gives you v3. And v3 is something that
38:15
I'm proud of because I've
38:17
taken the time to actually iterate with
38:19
it, tell it what it got
38:20
right, tell it what it got wrong.
38:22
And it knows I'm crystallizing my
38:24
information and my process.
38:25
That way I don't have to keep
38:26
re-explaining myself. And I think this is
38:28
the aha moment for people when they
38:30
realize that they can
38:30
go from zero to their finished product,
38:33
especially in the reporting
38:35
world where reporting takes
38:36
freaking hours. Like some people's jobs
38:39
is just reporting, right?
38:41
And fine, but I want to make
38:43
your life easier so that you can actually
38:44
go build more
38:45
relationships, focus on strategy,
38:48
be creative, or do quality assurance.
38:50
Those are the four main categories that
38:52
people should be spending
38:52
their time on. This is how you get that.
38:54
This is how you remove
38:55
copy and paste from your life.
38:57
So hopefully that's a little bit more of
38:59
a, like, take these steps to make a
39:02
really high quality skill.
39:03
This is what people pay us to do.
39:06
Love it. Well, Grant, thank you very much
39:07
for being on here, sharing
39:09
with us. If anybody wants
39:11
to connect with Grant, he is on all the
39:13
platforms, but I'm happy to
39:14
connect you with him as well.
39:15
You know where to find me, Andrew, at
39:17
foxwell.com. Until next time, though,
39:20
folks, keep on building.
39:22
Bye.
39:23
All right.
39:23
Bye.