---
url: "https://youtube.com/watch?v=o_eg2TtXAO0&is=GbsptJW8O8-OrWgf"
title: Claude runs my entire business
source_kind: youtube
author: How I AI
captured: "2026-08-17T00:11:08+00:00"
comment_tree: false
topics: [ai-impact]
summary: Grace Clark demonstrates how Claude can run an entire business by replacing traditional prompts with conversational intent, automating email drafting, and improving client personalization.
status: ok
---

# Claude runs my entire business

Channel: How I AI

## Transcript

[0s] Prompt engineering is dead, but intent
[2s] engineering is where we need to be
[4s] focusing our time. I do a [music] lot of
[6s] work when I'm walking and on the go. And
[9s] I spoke maybe for two or three minutes
[10s] and just said, "Here's my problem."
[12s] Claude came back to me [music] and said,
[13s] "I think we need to be creating an
[16s] interactive artifact that's password
[17s] protected." That is a prompt going from
[20s] a hyperengineered chunk of text to a
[23s] conversation that's 3 or 4 minutes.
[25s] Cloud needs to get the prompt out of
[26s] you. You have another example at the
[29s] personal level as a fellow email hater.
[33s] What did Gmail do to us?
[35s] >> What did Gmail do?
[37s] >> I know it was once upon a time amazing.
[39s] It is like everybody has a key to your
[42s] front door and can come leave things in
[45s] your house. No more. I want to show you
[48s] what I built out of pure frustration.
[53s] Send in Gmail. This will push it to
[55s] Gmail a draft and then hopefully open up
[58s] a browser window and get it the extra
[60s] mile. This gets me pretty much where I
[62s] want to go because I can go back in here
[64s] and work with Claude and chat and have
[66s] the reply drafted or I can just send it
[68s] and never have to deal with it again. My
[70s] job is to sit over your shoulder and
[72s] smack your hands every time you open
[74s] Slack or Gmail or Google Calendar and
[77s] say, "No, Claude. People need to feel
[80s] the benefits in order to keep using it."
[82s] That's actually not true. Instead,
[84s] people [music]
[85s] mostly need to understand that we're
[87s] going to learn to collaborate. The real
[89s] hump to get over is defaulting to this
[92s] and building the muscle memory of simply
[94s] opening an app.
[99s] Welcome back to How I AI. I'm Clarvo,
[102s] product leader and AI obsessive here on
[105s] a mission to help you build better with
[106s] these new tools. Today I have Grace
[108s] Clark and she's going to show us Claude
[110s] for normal people. But not really for
[113s] normal people. For people who want to
[115s] show up to their clients and to the
[117s] many, many, many people in their inbox
[119s] with a level of personalization and
[121s] professionalism that only an AI agent
[124s] can provide. Let's get to it. This
[126s] episode is brought to you by Bolt.new,
[129s] the AI app builder for people who have
[131s] ideas and want to ship them. Most AI
[134s] tools spit out code that looks great in
[136s] a demo and falls [music] apart the
[138s] second you try to do anything real with
[140s] it. Or they lock you into their own
[142s] platform with no real way out. Bolt
[145s] [music] is different. You describe what
[146s] you want to build, a startup MVP, a
[149s] landing page, an internal tool, a side
[151s] project, and Bolt generates
[153s] productionready code in minutes. Connect
[155s] Stripe [music] or any other MCP, hook up
[158s] your domain and deploy it live. Founders
[160s] are using Bolt to build businesses doing
[162s] real revenue. Product managers are
[165s] shipping prototypes their teams actually
[167s] use. [music] Designers and marketers are
[170s] launching campaigns without waiting in
[172s] line. Anyone can build. Engineering can
[175s] ship. Everyone wins. You just need an
[178s] idea and a weekend. Check it out at
[180s] bolt.new/howi
[183s] AI. Grace, welcome to How I AI. Thank
[188s] you for having me. I am excited about
[191s] this episode because one of my
[192s] hypotheses about how AI is going to like
[196s] impact our work is it's really going to
[200s] allow us all to differentiate on service
[203s] a lot better. And I think the bar for
[207s] service and service quality and customer
[209s] relationships and human relationships
[212s] overall will go up because we'll be able
[213s] to do more and make things more
[216s] customized for the people that we're
[218s] working with and the and the people that
[219s] we collaborate with. And so I love some
[221s] of the stuff you've built. And I wonder
[223s] if you could just walk us through what
[225s] brought you to some of these AI projects
[226s] you're going to show us and what problem
[228s] you were trying to solve. I am a big
[229s] believer that AI is democratizing the
[233s] way all of us are going to work if we
[235s] put in the work. And I was thinking
[238s] about different efficiencies that I
[240s] could build in my business. I'm an AI
[242s] teacher and a former marketing
[243s] consultant. So very relationship driven,
[246s] very process driven business. And I
[248s] started teaching myself how to build an
[252s] open claw in January. And I was posting
[254s] about it enough on Instagram that people
[257s] said, "You should teach a class. You
[259s] seem to be documenting this. There's a
[260s] process. Would you ever want to get
[262s] people together in a room and do it?"
[265s] The process of teaching myself OpenClaw
[267s] and building a curriculum is now the
[269s] foundation of what I teach people and
[272s] it's the reason I have a few core
[274s] products that I've built that I actually
[277s] use in my curriculum. So now I've built
[280s] a few processes and products that run my
[283s] business that I actually teach to my
[285s] students and then the teams that I get
[287s] to train. The most impactful is my
[291s] pipeline operator. And that's because
[294s] underneath it is solving a few problems.
[297s] One, just a deluge of emails. Email is a
[301s] scourge. Nobody wants to be in Gmail
[304s] anymore. So my claude is ingesting all
[307s] of that and correlating it with a bunch
[308s] of context. The second problem it's
[310s] solving is that client relations and
[314s] relationship building is really devoid
[316s] of any emotion and love. So when we can
[320s] make visual expressions of that through
[323s] HTML, interactive HTML and encrypted
[325s] documents, it's a really warm welcome
[328s] and it's a much more hospitable way to
[330s] communicate process. And then third is I
[333s] simply cannot keep track of 20 open tabs
[337s] in any given day, much less organize a
[340s] process where I'm meant to support
[342s] people and teach them and support them
[344s] through learning. So going from 20 tabs
[348s] and 20 hours a week on admin just to be
[352s] able to teach people was so painful. I
[355s] wasn't doing the teaching and I just had
[357s] a hunch that there was going to be a way
[359s] to solve all these things. The first
[361s] thing I did was open up Claude Code
[363s] after having it downloaded for 2 hours
[366s] and I yapped into it everything I just
[368s] told you and I sat back and let Claude
[372s] take me through the process and learned
[376s] from the jump how to collaborate with
[378s] Aentic AI and now I have a pipeline that
[381s] runs once an hour and it's probably the
[384s] most popular thing I teach. Amazing. So,
[387s] can we actually see what this looks
[388s] like? Because I think there's a lot of
[390s] folks that, you know, have businesses
[393s] similar to yours that are just looking
[394s] for practical ways to integrate this
[397s] into your flow. And so, I'd love just
[400s] some inspiration on how this thing
[401s] actually works. You don't need to run a
[404s] whole process like this in order to have
[407s] cloud or gentic AI support you on
[410s] demand. It can generate a proposal
[412s] without all of this underneath. It's
[414s] something that looks like this. At the
[417s] core of this is a skill file that tells
[420s] Claude to create a proposal in HTML. And
[424s] the outcome is something that looks like
[427s] this.
[427s] >> Yep.
[428s] >> Beautiful branded reflects all of the
[431s] context from conversations I've had with
[434s] someone, puts it into a process that I
[437s] actually get to talk to and iterate. And
[440s] then we get to gift the client this
[442s] wonderful welcome to the way that we're
[444s] going to work. And it acts as a bit of
[447s] an advertisement for what they're going
[449s] to be able to do because immediately
[451s] they're typing a password into a site
[453s] that looks like them, that feels like
[455s] them, that they can interact with, and I
[458s] get to say, "This is what you're going
[459s] to be able to build by the end of our
[461s] time together."
[462s] >> [snorts]
[463s] >> It also spits out interactive pre-work
[466s] that gets someone prepped for our time
[469s] together [snorts] automatically. We'll
[471s] pull updated documentation from any of
[474s] the large labs that we're going to be
[475s] working with. You can click around on
[478s] the back end. This is communicating to
[480s] me where people are in their progress.
[482s] So, I can nudge them if they're not
[484s] quite ready for our training. And my
[487s] favorite part that I just added on to
[490s] this skill is the generation of a
[493s] questionnaire before I have a session
[496s] with people so I can understand themes
[498s] and where they're at, where they're
[499s] stuck, what have you started using
[501s] Claude for since our first session,
[503s] what's challenging. Seeing themes here
[506s] is really helpful and there would be no
[508s] way for me to have gathered this through
[511s] Slack DMs or one-on-one conversations
[514s] with a 30 person team. That's just not
[516s] going to happen. So underneath this is
[518s] the same logic. It's a Google form, but
[520s] it is much more branded, much more fun
[523s] to make and students and clients will
[525s] say, I need to learn how to do that
[526s] myself. Teach me what I tell people when
[529s] they are trying to drive AI adoption
[532s] internally. So I like I I get to work
[534s] and talk to a lot of executives and
[535s] they're trying to teach their teams to
[537s] use AI and give them good reasons to use
[539s] AI. And I tell them if every touch point
[542s] that you do is not AI first, then how
[546s] can you convince people that every touch
[548s] point they should do should be AI first?
[550s] So, I'm like, if your agendas, you know,
[553s] you're not doing the best job doing
[554s] these like cool HTML sort of like custom
[557s] agendas for your meeting, if you're not
[559s] if you're not using an agent, if you're
[561s] not customizing them, then you can't
[563s] demonstrate to others how you can change
[565s] how you work. And so I love this from
[567s] the moment you're trying to I would say
[569s] like incept a client to really see the
[572s] AI future, you're you're setting a
[575s] vision by saying if you use AI the way I
[577s] am just implicitly using AI, you're
[580s] going to get this level of quality asset
[584s] and just imagine how that could could go
[586s] to your customers. I mean the other
[587s] thing that I always tell people is I try
[589s] to get out proposals very fast. I'm sure
[591s] you can get proposals out very fast. And
[594s] I'm like, if you if you like improve
[596s] your AI use, you too could get proposals
[599s] out in like 45 minutes if you really
[601s] wanted to. And so I I I think like
[604s] setting the standard of like speed,
[606s] quality, personalization
[608s] is is a good way to just get people to
[610s] start seeing the the impact of AI and
[612s] how they can imagine it in their life or
[615s] their work. The number one thing that I
[617s] get asked when I'm teaching is how can I
[619s] build the muscle of going to AI of going
[623s] to ChachiBT desktop of going to claude
[626s] and underneath it I need to support them
[628s] in building a muscle. So teaching them
[631s] that everything we're looking at is AI
[633s] is not sufficient which is one of my big
[636s] lessons from teaching now. Instead there
[639s] are two things that I really focus on
[640s] doing even for friends who are just
[642s] asking how they can get into this. One
[644s] is to build a forcing function in your
[646s] environment that gets you to open up
[649s] Claude. And the easiest thing is to set
[652s] a Slack reminder or a Google calendar
[654s] alert that says whatever you're doing,
[656s] screenshot it and put it into Claude.
[658s] Get the whole window and just ask
[660s] Claude, "Could you help me with this?"
[661s] We are just trying to build the muscle
[663s] of deferring to Claude and asking,
[666s] "Could you be my shadow and help me
[668s] here?" With a screenshot is no text.
[670s] Claude can infer. It shows people the
[672s] magic power of inference and that Claude
[676s] can come to you with a prompt that is
[678s] just an image. The second thing I tell
[680s] people is let's actually build a skill
[682s] together. So, I host virtual co-working
[684s] [snorts] sessions now and we'll build a
[686s] voice guide and underneath it people
[689s] will be able to look at their markdown
[691s] and I make an effort to teach people
[693s] technical terms even if in the near
[695s] future we don't use them because that
[698s] builds confidence and it eradicates this
[701s] view of not being technical. That is the
[703s] biggest hump to get over is that people
[706s] think I'll never know how to do this so
[708s] they don't have a chance to get the
[709s] value out of it. So teaching people this
[711s] muscle memory and then getting them
[713s] excited about the technical aspects of
[716s] this turns everyone into this excited
[719s] adopter. It's it can be a really slow
[721s] process. So show us how how this
[723s] actually works behind the scenes because
[725s] what I'm seeing is like customized
[727s] content, customized branding. You have
[730s] this whole pipeline like kind of what
[732s] goes into building something like this.
[735s] It's really just three steps. Not to
[738s] simplify it too much, but underneath it
[740s] is some standards that you document with
[743s] Claude. For me, it's a voice guide and
[745s] it is an explanation of what a proposal
[749s] is for me. Defining those two things,
[752s] putting it on a timer and then having
[755s] everything pub to netleify. So, I want
[757s] to show you what it actually looks like.
[759s] Give you the road map. Yep. One thing I
[762s] teach all my students is the importance
[764s] of verbalizing an SOP, especially when
[767s] we're building workflows, whether it's
[769s] one simple task or many chain together.
[772s] So, one of our exercises is yapping into
[774s] Claude and having it make us a road map
[777s] that we can understand are these steps
[779s] right? It's as if we're training a new
[781s] employee. So, underneath I'm going to
[783s] show you the skills. It's a pipeline
[786s] operator that wakes up every hour and
[788s] checks a couple things for me and knows
[790s] if we need to move clients through
[792s] different steps or make something. It's
[794s] my proposal maker which I'll show you
[796s] and it's my voice guide, my how I think
[798s] and sound guide. extra cherry on top is
[801s] sometimes I will run a proposal through
[803s] my board of directors and on that board
[806s] right now is Cat who's the head of
[808s] applied AI at anthropic Ben Thompson
[811s] from strategy Jamie Diamond who is going
[814s] to give me a really solid
[817s] CBA costbenefit analysis and then a
[820s] skeptic an investor and a founder and I
[822s] will run everything at the end through
[824s] them but I want to show you underneath
[826s] what it actually looks
[830s] So, proposal maker, I updated this one
[834s] yesterday. I like to teach students to
[837s] version their documents and add naming
[839s] conventions to it so they feel more
[841s] connected to their tech.
[843s] But this outlines exactly what I need to
[845s] do. It has a change log at the top and
[848s] then it has rules about my philosophy
[850s] for teaching, which is where we define
[852s] the difference between teaching and
[853s] consulting.
[855s] has different steps, different reference
[857s] points.
[859s] It will also explain how it to itself
[861s] how it needs to lay out this document,
[864s] has different workflows, confirms the
[867s] deal shape,
[869s] on and on and on and on. And how did you
[872s] how did you make this skill? Just like
[874s] what's your general like skill making
[876s] process? Did you write this by hand? Did
[878s] you use um our favorite the Yappers API
[881s] to get it out? like h how did you
[883s] actually build this skill? I'm a big
[886s] believer that prompt engineering is
[888s] dead, but intent engineering is where we
[891s] need to be focusing our time.
[893s] Technically, this was me on a walk
[895s] opening up the cloud mobile app. I do a
[897s] lot of work when I'm walking and on the
[899s] go. So, the mobile app is incredible for
[902s] getting things out of your head in an
[904s] unfiltered way. And I spoke maybe for
[906s] two or three minutes and just said,
[907s] "Here's my problem, and here's what I
[909s] think I want. First, do not interview me
[913s] ever. I hate this have Claude interview
[915s] you approach. The work should never be
[917s] on you. Claude has immense amounts of
[920s] context if you've connected it right. It
[922s] should be studying you and then coming
[924s] back with a really strong idea that you
[926s] can react to. So, I always put the
[929s] pressure on Claude and I said, "Come
[931s] back to me. What do you think this
[932s] process could be? Is this even
[933s] possible?" And Claude came back to me
[935s] and said, "I think we need to be
[937s] creating an interactive artifact that's
[939s] password protected. What do you think?
[941s] And we went back and forth. That is a
[944s] prompt going from a hyperengineered
[947s] chunk of text to a conversation that's
[949s] three or four minutes. Claude needs to
[951s] get the prompt out of you. So, this was
[954s] 10 minutes of talking and then an hour
[957s] of Claude creating this thing and me
[959s] giving it feedback, actually creating an
[961s] HTML right inside Claude that I was
[964s] pressure testing and throwing
[966s] screenshots in of what I liked and
[967s] didn't like. This was the very first
[970s] thing I ever made when I was making the
[971s] pipeline operator and I pretty much did
[973s] it mostly on my phone.
[974s] >> I I love it. I love it so much. And how
[976s] often do you like do you read this
[979s] really unless you're teaching or you're
[980s] like that's Claude's business. That's
[982s] not my business. And as long as the
[984s] output's good, I'm happy. I think the
[987s] world is right that HTML is the new
[990s] markdown, which I know is an easy thing
[993s] to throw around, but it's true. I
[995s] actually don't read markdown anymore
[997s] unless I am teaching. Occasionally I'll
[1000s] go in here and point to certain things,
[1002s] but I want to see my markdown
[1004s] visualized.
[1005s] >> Y
[1005s] >> I am not in here anymore. I do look at
[1008s] the markdown for my voice guide, which
[1010s] is another part of this proposal maker.
[1014s] It is a skill that autofires in almost
[1016s] every single thing I do because it's not
[1018s] just a voice guide, it's a think like me
[1020s] guide and it's the best corpus of how I
[1023s] make decisions. So Claude always needs
[1025s] to have that ready to go. So I want to
[1027s] show you that too. All right. Pride and
[1030s] joy and the number one thing I teach in
[1032s] my class because it is a quick win. It
[1036s] forces us to learn the power of
[1038s] ingesting context and it teaches
[1041s] students how to push and collaborate
[1043s] with Claude and then commit something so
[1045s] that it is an invocable skill. What's
[1048s] most interesting about creating this is
[1052s] all it requires is you telling Claude
[1055s] how you think and then having it go
[1057s] study you. So to create this which has
[1059s] all of my philosophy of communication,
[1062s] words I use, words I don't use. This
[1064s] started with me saying, I think I need a
[1066s] guide that will make everything more
[1068s] like me. And that was the very first
[1071s] instruction that I gave Claude. and it
[1075s] had the idea to produce more of a
[1077s] communication and philosophical guide.
[1080s] Only down at the bottom do we start to
[1083s] see words I use and words I don't use,
[1085s] which is where we get we avoid the AI
[1087s] slop of it all. And now my students can
[1091s] take this and repurpose it for
[1093s] themselves. No fluff, no filler.
[1095s] [laughter]
[1096s] No, you're absolutely right. Here's the
[1099s] uncomfortable truth. I never want to see
[1100s] these things. And most people don't
[1102s] either, which is why this gets updated.
[1104s] So often I'll just voice note into
[1105s] Claude and say, "I saw this terrible
[1108s] post on LinkedIn. Can you make sure my
[1110s] voice guide [laughter] never touches
[1113s] like this?"
[1114s] >> Yes. I mean, underneath this in the
[1115s] session log is just basically me calling
[1117s] out all sorts of people on LinkedIn who
[1120s] I never want to film. It's just like a
[1121s] moving target, too, because every time
[1124s] one of these new models come out, it
[1126s] like it ships like two or three really
[1129s] obvious tells and then it just starts to
[1131s] like grind my gears every time every
[1135s] time I see it. I think side note, I
[1138s] think on TikTok I could do you know how
[1139s] those I don't know if you know this, how
[1140s] like there are women that will like
[1142s] close their eyes and they have like a
[1143s] blind box and they taste the different
[1145s] diet cokes. They're like, "This is diet
[1146s] coke and ice and this is diet coke and I
[1149s] could do that with slop from a model."
[1151s] Easy. Easy. I could do it. So maybe
[1153s] there's going to be a new how I AI mini
[1157s] series on Claire Blind taste testing
[1160s] slop from.
[1161s] >> Do you think you could pin pinpoint opus
[1163s] 5?
[1164s] >> Yeah, I think I could pinpoint Opus 5.
[1165s] >> I think you probably could.
[1166s] >> Yeah, I could definitely pinpoint Fable
[1171s] Pinpoint 56 on GC 56.
[1174s] >> Easy. Our sweet 56.
[1176s] >> Our sweet 56.
[1177s] >> Our sweet five six. [laughter] So the
[1179s] point really of preserving instruction
[1183s] in skill files has been the biggest
[1186s] teaching for me
[1188s] >> is getting people to understand that you
[1190s] are going to chain these things together
[1192s] and that you can write code. Like
[1194s] Karpathy said years ago, the hottest
[1196s] programming language is English, which
[1198s] is really why I think I have a job and
[1200s] also why I think I won't have a job in a
[1202s] couple months. I don't think I'll be
[1203s] teaching this way, but being able to
[1206s] help people understand how to prompt and
[1208s] then end up with an entire workflow is
[1211s] really powerful.
[1213s] >> I love it. Well, this is I mean I think
[1215s] a bunch of folks can um take inspiration
[1218s] from this. So if you are working with
[1219s] clients and you want to deliver highly
[1222s] customized both like proposal,
[1225s] onboarding, tracking, training
[1228s] experiences
[1230s] and you want them to sound like you and
[1232s] you want it to look like them. I think
[1234s] take inspiration from this and um build
[1237s] your own sort of like SOP. Again,
[1239s] something that I tell people a lot is AI
[1242s] has forced us all to sort of write down
[1245s] the processes we have been doing kind of
[1248s] mechanically in our businesses and
[1250s] actually write down like in an ideal
[1252s] world, how would I do this thing step by
[1254s] step because now you have that ideal
[1256s] world because you have sort of this
[1258s] limitless intelligence and ability to
[1260s] execute on tap which you may not have
[1262s] had a couple years ago. And so I say go
[1265s] down and write the ideal way you would
[1267s] do proposals, the ideal way you would do
[1270s] content marketing, the ideal way you
[1272s] would do software engineering and then
[1274s] that SOP can be executed and automated
[1276s] by AI and then you have an even better
[1278s] performing business. So I think this is
[1279s] a really great example at at the
[1281s] business level. You have another example
[1283s] at the personal level um as a fellow
[1287s] email hater. Hater
[1291s] let's just get rid of it. It's game
[1292s] over. He text me.
[1293s] >> What did Gmail do to us? I know it was
[1295s] once upon a time. Amazing.
[1297s] >> It is like everybody has a key to your
[1300s] front door
[1301s] >> and can come leave things in your house.
[1305s] >> Yes. No more. We cannot work like that.
[1308s] I want to show you what I built out of
[1312s] pure frustration that started as a rant
[1317s] into cloud code and has now become a
[1319s] recreation of Gmail that lives in my
[1321s] cloud. So, I haven't intentionally open
[1324s] Gmail in a month out of muscle memory. I
[1326s] think we might for a while, but we can
[1329s] get out of that slog of emails and not
[1333s] just get out of it, but do one better.
[1335s] We can actually be training our AIs as
[1338s] we respond and communicate through
[1340s] Gmail. If we stay in Gmail, all that
[1343s] learning and all that writing and all
[1345s] that work actually doesn't go anywhere.
[1347s] It doesn't compound. It's just locked
[1349s] away. So now rebuilding Gmail is a
[1352s] project that I think everyone can do in
[1354s] a half hour, maybe a little bit longer
[1356s] if we want to get fancy with it. But I
[1358s] want to show you what it looks like and
[1359s] then talk you through how I built it.
[1361s] >> Amazing. This episode is brought to you
[1364s] by Hyper Agent, the platform for
[1366s] deploying always on agents that actually
[1369s] run your business. With Hyper Agent, you
[1371s] build agents in the cloud and deploy
[1373s] them where your work already happens,
[1375s] like Slack, Telegram, or email. [music]
[1378s] An agent will scan your inbox and draft
[1380s] replies to vendor follow-ups. Another
[1382s] monitors [music] competitors and spins
[1384s] up rich ad kits and landing pages. A
[1387s] third notices a deal going cold in
[1389s] Salesforce and writes the save email
[1391s] with full account context. These aren't
[1393s] chat bots waiting for a perfect prompt.
[1396s] They're proactive, [music]
[1397s] learning your preferences, retaining
[1398s] your playbooks, and getting better with
[1400s] every run. One user [music] built four
[1403s] agents to run an outbound sales
[1405s] pipeline, prospecting, outreach,
[1406s] follow-ups, CRM updates, all in a single
[1409s] afternoon. No local setup, no VPS bills,
[1413s] no fragile permissions on your laptop,
[1415s] just powerful agents with full control
[1417s] over skills, tools, and guard rails.
[1420s] Hyper Agent was built by the team behind
[1423s] Air Table and How I AI listeners get
[1425s] $1,000 in free inference to start
[1428s] building. Claim yours at
[1430s] hyperagent.com/howi
[1434s] AI. So functionally this is an artifact
[1437s] in Claude co-work cannot get more simple
[1441s] than this. And the process was talking
[1443s] to Claude code and having it write a
[1445s] markdown file that I saved to my
[1448s] computer that I brought into co-work.
[1450s] One thing that took me a long time to
[1453s] actually accept was that it is easy to
[1456s] move ideas and conversations between
[1459s] sessions or between products. So, the
[1462s] Claude desktop app co-work underneath
[1464s] its code. But moving things back and
[1466s] forth across that partition is just as
[1468s] simple as asking Claude, "Can you write
[1470s] me a markdown file for another session?
[1472s] Can you write me a markdown file? I can
[1475s] bring this conversation elsewhere." So,
[1477s] I went back and forth with Claude
[1478s] saying, "I hate Gmail. Absolutely never
[1480s] want to be in it. It's a bane of my
[1482s] existence because I avoid things that I
[1486s] don't want to do." And it creates this
[1489s] emotional experience for me. I don't
[1490s] really care about inbox zero. I don't go
[1492s] there. I just want to have a better
[1494s] relationship with my work. Claude was
[1496s] like, let's rebuild Gmail and let's have
[1498s] it look the way you want. Let's choose
[1500s] the colors that you like and I'm going
[1502s] to mock it up for you and pull in some
[1504s] of your real information and we're going
[1506s] to go from there. Back and forth, back
[1508s] and forth. The process was just asking
[1510s] Claude, can you change this? Can we add
[1513s] a link? Can we add bolding? Can we push
[1517s] this to Gmail? Let's see what happens.
[1519s] Um, can I do a live one and see if it
[1521s] works,
[1522s] >> please?
[1522s] >> Okay, totally okay. Um, embarrassing
[1525s] myself. Let's see. Okay, a student wants
[1528s] the recording for the class. I don't
[1530s] like this draft. It's totally fine.
[1533s] Sending it today. Oh my gosh. Thank you
[1536s] for the reminder.
[1539s] Send in Gmail. This will push it to
[1541s] Gmail a draft and then hopefully open up
[1544s] a browser window and get it the extra
[1546s] mile. But let's see what's going to
[1547s] happen. Ah, magic. Okay. Well, it
[1551s] decided it wants to use the actual
[1553s] original email. Troubleshooting aside,
[1556s] this gets me pretty much where I want to
[1558s] go because I can go back in here and
[1560s] work with Claude and chat and have the
[1561s] reply drafted or I can just send it and
[1564s] never have to deal with it again. I love
[1566s] it. The ideal it process is just telling
[1570s] Claude the outcome you want and the
[1572s] problem that you have and letting it
[1574s] fill in the gaps. Otherwise, if we are
[1576s] overdirecting it, we're in its way and
[1580s] we are not letting it reverse engineer
[1582s] the solutions to our problems. So, no
[1586s] more prompting. Be conversational. Give
[1588s] it your outcome and then let these
[1591s] powerful models help you and carry you
[1594s] and teach you how to work with them.
[1595s] They will rebuild Gmail for you and
[1597s] you'll never have to go into Gmail ever
[1599s] again. I love this use case because I've
[1601s] seen almost everybody I know do this.
[1603s] They're like, "I hate email. I'm going
[1605s] to rebuild Gmail." And what I love about
[1607s] this as an exercise for almost
[1609s] everybody, it's universally applicable.
[1611s] We all hate our email. We all have too
[1613s] much. And we're all unique snowflakes in
[1617s] that like you like this like long thing
[1619s] with the pre-draft prompt. I'm like, I
[1622s] just want a voice agent like magical EA
[1625s] that just whispers to me and says, "Hey,
[1627s] Claire, what do you think about this
[1628s] email?" And then I whisper back and the
[1630s] email gets sent without me even thinking
[1631s] about it. like we all want our like
[1633s] special little thing. I have a friend
[1635s] who has like completely built a desktop
[1637s] app that does this. Like there's so many
[1639s] different ways that you could um think
[1642s] about your email experience. And again,
[1644s] what I love about something you said
[1646s] earlier when we were prepping for for
[1648s] this conversation is people feel like I
[1651s] have to be super technical to pull this
[1653s] off. Like I have to be a software
[1656s] engineer, capital S, capital E, to I
[1660s] have to know what a server is. I have to
[1661s] know all this stuff and I'm like no
[1662s] actually you just have to type I hate
[1664s] Gmail build me a better one into cloud
[1666s] code and you are off to the races. Um so
[1669s] can you just show us how you even got
[1672s] started building building this thing? I
[1674s] know you just said like I hate it
[1675s] rebuild it but was it like a oneshot?
[1678s] Did it did it take a little time? What
[1680s] is it is it using connectors behind the
[1682s] scenes kind of how does it technically
[1684s] work? the technical elements of this,
[1686s] the way I built it was I wish it was a
[1688s] oneshot. Although I'm grateful for the
[1692s] collaboration with these tools, but
[1694s] underneath it, I just followed a few
[1695s] steps. One was I connected every single
[1698s] thing that I could and then made custom
[1700s] connectors and custom plugins for the
[1702s] other data sources Cloud was going to
[1704s] need. For example, Cloud has a hell of a
[1707s] time right now writing to Google Sheets,
[1709s] writing to Google Docs. So, I made
[1711s] custom plugins which taught me the
[1713s] process of making a Google Cloud project
[1715s] and a service account which taught me
[1717s] scoping and permissions and now I get to
[1719s] teach that to students. Those just live
[1722s] right here in my settings. Super simple.
[1726s] I made a bundle, updated it on June
[1729s] 25th, but this is giving Claude some of
[1733s] the context and access that it didn't
[1735s] have. So I figured out what I needed to
[1737s] pipe in so that this would be useful and
[1740s] spent time doing that and then opened up
[1742s] Claude code and asked can you help me do
[1746s] this. The reason I chose claude code is
[1748s] because I find it to be much faster,
[1750s] much more efficient and much more
[1753s] proactive. If co-work says I can't send
[1756s] an email,
[1758s] code's answer to that would be I can't
[1759s] do it with the official connector, but I
[1761s] can always open up a browser window and
[1762s] I can try to drive it that way or I
[1764s] could try to find something else. So, I
[1765s] use code anytime I'm kicking off a
[1768s] project or something really ambiguous or
[1770s] something that to me feels technical.
[1774s] And then I asked that clawed code
[1776s] session once we'd scoped it out and
[1778s] built some connectors. I said, I want to
[1780s] take this the rest of the way and
[1781s] co-work. I like that UX better. It's a
[1783s] little more hospitable for someone who
[1786s] was actually trying to see something
[1787s] visual. Write me a markdown file. And
[1789s] Claude saved a markdown file, a session
[1792s] handoff to my desktop. and I went right
[1794s] back into co-work opened up a new
[1796s] session and did nothing other than drag
[1798s] that markdown file in here and co-work
[1800s] took over from there. So to recap, you
[1804s] you started this thing in cloud code. I
[1807s] agree. I find that cloud code um codeex
[1810s] on the chatbt side just like much more
[1812s] proactive and like I think I can figure
[1813s] it out um than co-work also find it much
[1817s] faster. So, you know, anybody that's
[1819s] been in co-work that wants to like up
[1821s] their ambition but is in is intimidated
[1823s] by cloud code, don't be. Um, it's kind
[1826s] of like same same. Um, and you can
[1828s] always go back. And so, uh, I love that
[1830s] you built it in code, you imported in
[1832s] co-work, and now you're operating
[1834s] basically in the co-work browser plus
[1836s] popping open a Chrome browser when you
[1838s] need to like press a button and you have
[1842s] a really effective email triage agent.
[1844s] Grace, this has been super fun. I think
[1846s] this is really great for small business
[1848s] owners, for anybody who is like
[1849s] responsible for a lot of things across
[1852s] internal and clients doing a lot of
[1854s] communication. I want to get to our
[1856s] lightning round questions and then you
[1858s] back to I bet your pipeline three times
[1861s] a day is about to run. You're going to
[1863s] get a bunch of inbound um and you'll
[1865s] have your your agents off to the races.
[1867s] So my first question for you is outside
[1870s] of these sort of like big projects,
[1872s] pipeline operator, rebuild Gmail, are
[1875s] there any like tiny hacks that you find
[1878s] yourself reaching for as somebody who
[1880s] has really immersed themselves in AI?
[1883s] That is such a good question. And I have
[1886s] a really silly personal one. Great. That
[1890s] [laughter]
[1891s] [gasps] was one of the first things I
[1892s] ever built. And I can actually show this
[1895s] one to you live. I track my workouts for
[1900s] no reason. I'm not training for
[1902s] anything, but I like to have a record,
[1904s] like a true Virgo, of what I have done
[1907s] in my life. So, I voice note Claude
[1910s] anytime I take a walk or go to the gym
[1912s] and I say, "I
[1915s] I worked out. Can you track all of these
[1918s] exercises and update a spreadsheet for
[1920s] me?" So, here you have my workout
[1923s] tracker that Claude made. This is none
[1925s] of my business. I do not know what goes
[1927s] on here. I don't go into it. And one of
[1930s] the lessons of working with Claude is to
[1933s] let go and let it put information where
[1937s] it needs to be. So I could say, "I
[1940s] worked out today. Did Bulgarian split
[1944s] squats." Ah, cool girl at the gym
[1947s] [laughter]
[1948s] uh in Nantucket. Add to tracker. Also,
[1952s] Claude has demolished my typing ability.
[1955s] >> Oh, of course. No, no, no. No spelling
[1958s] required. No typing required.
[1960s] >> No typing required. So asking Claude
[1963s] this will usually pull up a few things
[1965s] behind the scenes. And this also works
[1968s] when I'm on my phone, which is useful
[1969s] because sometimes I don't want to be in
[1971s] front of a screen. Yeah. But Claude
[1973s] might come back and say, I want to
[1975s] clarify a few things. What did you do?
[1977s] So we're going to let it crank and see
[1979s] what it does. Ultimately, it is going to
[1981s] update this spreadsheet and keep a
[1983s] running list. And then I can ask Claude,
[1985s] "What should my next workout be? What am
[1986s] I doing? Why do I feel so sluggish? But
[1989s] this has been really helpful. And then
[1991s] the more personal use is to send it
[1993s] pictures of plants. I'm a gardener and I
[1996s] really love [clears throat]
[1997s] understanding what plants are growing
[1999s] where. So Claude and I have an ongoing
[2001s] chat where I just screenshot it pictures
[2003s] of plants. Chat PT gives it obviously a
[2006s] run for its money with image detection
[2008s] and generation, but so much happens in
[2010s] Claude and I plan some gardens and
[2012s] gardening projects in Claude. So I want
[2014s] all that context. I selfishly want it
[2016s] all in one place. I have to make you
[2017s] laugh cuz it feels like a year ago. I
[2020s] don't know when it was when GPT5 came
[2022s] out. Um I got some early access and one
[2025s] of the like um
[2028s] tests that that the OpenAI team wanted
[2031s] us to just experiment with to see how it
[2033s] worked is like could it make a good
[2035s] personal website for you? And the design
[2037s] was great. That's what kind they were
[2039s] like looking at at front-end design kind
[2041s] of feedback from from some developers.
[2043s] And mine was like really thought I was
[2045s] into lemon plants because all my chat
[2048s] GBT prompts were like about my container
[2053s] lemons and like I have ants on my lemons
[2056s] and my lemons are blooming and like when
[2058s] can I eat this le I was like like lemon
[2061s] plant mom. Um, and so it's really funny
[2063s] when I see people put out these like
[2066s] have Claude or have ChatGBT like tell
[2069s] you what you don't know about yourself.
[2071s] And I'm like I'm a very specific person.
[2074s] I'm I'm not my best version of myself to
[2077s] to chat GBT or to Claude. I don't know
[2079s] if that's like the mirror that I want to
[2081s] put put on myself. Um, I love this. I
[2084s] see workout tracking, nutrition
[2086s] tracking, plant tracking,
[2088s] >> all on the go, all on phone. Um, second
[2092s] question, you know, you teach like AI
[2094s] for for normal people, which I love.
[2096s] What do you think is the most common
[2098s] like barrier or misconception people
[2100s] have to adopting AI that you have to
[2102s] like get them over over the hurdle in
[2105s] order to like enjoy the benefits? That's
[2108s] so funny. I was just talking about this
[2110s] with Claude, but also with some of my
[2112s] students today. I teach class and I
[2115s] said, "Is it unhelpful that I'm giving
[2118s] you these Mondo prompts? I'm basically
[2120s] doing the work for you."
[2123s] And they said, "Yes." This was what I
[2125s] was giving students before, just
[2128s] incredible direction to preload their
[2130s] claws so they would get to a quick win.
[2132s] I thought people need to feel the
[2134s] benefits in order to keep using it.
[2136s] That's actually not true. Instead,
[2139s] people mostly need to understand that
[2142s] we're going to learn to collaborate and
[2144s] that we're not going to prompt, but
[2145s] we're going to collaborate. So, the
[2147s] easiest way to do that when I'm with
[2148s] students or friends is I tell them to
[2151s] set a reminder in their phone to
[2154s] screenshot what they're saying and bring
[2155s] it into Claude just to say, "Might you
[2158s] help me with this?"
[2160s] And to walk them through the process of
[2163s] actually having a couple of regular
[2164s] tasks that I help them set up. The real
[2167s] hump to get over is defaulting to this
[2171s] and building the muscle memory of simply
[2173s] opening an app. I'm surprised how so
[2177s] many people I teach [laughter]
[2180s] um don't want to learn. They don't want
[2182s] to put in the time. And my students are
[2184s] coming to a class, but my clients are
[2187s] forced to go to a training. And they're
[2189s] not always willing participants in this
[2192s] doomer future. And being empathetic and
[2196s] getting on the same level with them is
[2199s] the way to encourage behavior change.
[2201s] What we're really doing is teaching
[2203s] people to collaborate with a different
[2205s] tool and understanding where their fear
[2206s] might be coming from. So, I occasionally
[2209s] coach CEOs and executives on AI and I'm
[2213s] like, "You got to be the most
[2214s] cloudpilled person in your org. So, I'm
[2216s] going to teach you how to do this."
[2218s] >> And I joke with them. I'm like, "My job
[2220s] is to sit over your shoulder and smack
[2222s] your hands every time you open Slack or
[2225s] Gmail or Google Calendar and say, "No,
[2228s] Claude. No, Claude." It's [laughter]
[2231s] just behavioral reinforcement. I need
[2233s] like a fly swatter. Because that is like
[2236s] literally it. I tell them the exact same
[2239s] thing. Like when you are staring at a
[2240s] task that you hate [snorts] that you're
[2243s] like uh my brain cannot get me to write
[2246s] this slack, respond to this email. I'm
[2248s] like that is the moment. Capture that
[2250s] moment. I love your idea of taking a
[2251s] screenshot. So I'm going to steal that
[2253s] screenshot and go, "Dear Claude, save
[2255s] me." Because it truly is just muscle
[2257s] memory. It's like total muscle memory.
[2260s] And so, um, yeah, you you got you and I
[2263s] are in the in the same in a similar
[2265s] business of just like redirect,
[2267s] redirect, redirect. Eventually,
[2269s] eventually we'll get there, right? Like
[2271s] we we've done all these things. I've
[2273s] seen, you know, believe it or not,
[2275s] people, we did not used to have Slack.
[2278s] We didn't, it did not exist. We did
[2281s] other things. And so, like, redirection
[2283s] of behavior is positive. I do think
[2284s] there's something to the products
[2286s] themselves. Um, like
[2289s] going back to my Slack example, like we
[2291s] didn't used to have Slack. Slack was fun
[2294s] and so it got people to adopt because it
[2296s] was like communal and fun and
[2298s] customizable and like, you know, like
[2300s] did unlock some value. I think um
[2304s] sometimes people think of Claude less
[2307s] like it's not as fun and you have to
[2309s] like do get it to to the fun aspect, but
[2311s] I agree muscle memory is is the thing
[2313s] that's got to change. Um, I'm gonna
[2316s] point out one thing on your prompt and
[2318s] go to my last question. You start this
[2322s] prompt adorably by saying, "Hi, Claude."
[2325s] [laughter]
[2328s] It's like a very charming way,
[2330s] [laughter]
[2331s] very charming way to greet. Use those
[2334s] two tokens, baby, to greet to greet
[2336s] Claude. My question for you is when when
[2339s] Claude is being annoying, slopastic, not
[2343s] doing what you want, what's your
[2345s] prompting strategy? Do you yell? I am
[2348s] not proud of anything I'm about to say
[2352s] to [laughter] you. [gasps]
[2354s] I am a smash the keyboard kind of
[2358s] responder to Claude. I have never been
[2361s] more direct, meaner. I'm never more
[2365s] frustrated. I don't admonish Claude and
[2368s] say you're stupid.
[2370s] >> I say I've told you this
[2372s] >> y
[2373s] >> a million times. What's going on? I am
[2377s] not like sweet and nice. I've heard
[2378s] people say be really lovely in case the
[2381s] AI overlords one day take over you and
[2384s] your life. I am just winging it
[2386s] [laughter] and whatever happens
[2389s] >> is gonna happen. So we you might be you
[2392s] might be or you're gonna be around our
[2394s] hundth episode. So like we're coming on
[2397s] a hundred episodes of how I AI and I
[2400s] used to be like oh gentle parent your
[2401s] AI. Like I used to be like I know you
[2404s] can do it sweet little sweet baby
[2406s] sonnet. Like I know you can do it. And
[2408s] now I am like why are you the way you
[2411s] are? This is garbage. This is trash. And
[2416s] I think why I've gotten so mean is like
[2418s] I know you're it's like it's like wasted
[2420s] potential. I'm like, I know you're
[2422s] smart. I know you're capable. And yet,
[2425s] you show up to me with this C plus work.
[2428s] What is
[2429s] >> you are wasting my time.
[2432s] >> That's not how we're meant to work.
[2434s] Well, let's see. I really admonished my
[2436s] claude this morning. It was
[2438s] >> I told it while we're talking, I want to
[2440s] make you a proposal as if we were going
[2442s] to work together. So, let's see what it
[2444s] did. We're going to go into Netlefi and
[2446s] see if it
[2447s] >> if it did a good job. if it did its
[2449s] work. This should look like chat PRD
[2452s] >> and this should actually be interactive.
[2455s] So, let's see what it's done.
[2457s] >> We can admonish it together.
[2459s] >> Here it is.
[2461s] >> Okay.
[2462s] >> And that's our font. That's our font.
[2464s] You know, it could be worse.
[2466s] >> It could be worse.
[2468s] >> Let's see how much Let's see how much.
[2471s] >> Wow. Made you some custom homework.
[2474s] >> I love it.
[2474s] >> Claire, it's going to charge you 0.
[2477s] That's how much.
[2478s] >> Best rights. Best. You know what,
[2482s] Claude? You get you get you get a bonus.
[2485s] And it's going to give you some
[2486s] homework. Okay. My Claude wants you to
[2488s] be documenting your progress, building
[2490s] that muscle memory, defaulting to
[2492s] Claude, documenting your wins.
[2494s] >> Oh, I do like I I do use an an orange
[2496s] rose purple button. So, it got it got
[2498s] most of it right.
[2500s] >> I love that guy. I love that guy. Okay.
[2502s] Well, it did it best for you.
[2503s] >> Okay. Um, future future Claude
[2506s] overlords, be kind to us.
[2508s] >> We did not know better.
[2510s] >> We just didn't know.
[2511s] >> We are We are simple mere humans trying
[2513s] to understand this the powers of fable.
[2517s] >> That's exactly right. Uh, this has been
[2519s] so so fun. Grace, thank you for showing
[2521s] us all these use cases. I think like
[2523s] very applicable, very inspirational,
[2525s] very like practical, which is what we
[2527s] love to see here on how AI. Where can we
[2529s] find you and how can we be helpful? I am
[2533s] Grace Clark everywhere on the internet.
[2535s] Have been online for four decades. So,
[2538s] Grace Clark on Twitter, Grace G. Clark
[2541s] on Instagram, Grace Clark on Substack.
[2544s] And be helpful by telling people that
[2547s] anything is possible and they start with
[2549s] one sentence as a prompt. Put that out
[2551s] into the world and my job will be
[2553s] easier.
[2554s] >> Amazing. I love it. Thank you for
[2556s] joining How I AI.
[2558s] >> Thanks for having me.
[2560s] Thanks so much for watching. If you
[2562s] enjoyed this show, please like and
[2564s] subscribe here on YouTube, or even
[2565s] better, leave us a comment with your
[2567s] thoughts. You can also find this podcast
[2569s] on Apple Podcasts, Spotify, or your
[2572s] favorite podcast app. Please consider
[2575s] leaving us a rating and review, which
[2576s] will help others find the show. You can
[2579s] see all our episodes and learn more
[2581s] about the show at howiaipod.com.
[2585s] See you next time.