Every's creative lead Lucas Crespo and Kieran Klaassen attempt to build an AI design pipeline that captures Lucas' artistic style and automates mechanical parts of banner creation, using Midjourney and browser automation to replace repetitive prompt-tweaking. # Channel Identity Observations — "Can AI Learn our Designer's Eye?" ## Observable Format & Presentation - **Live-streamed, conversational demo format**: The video opens with a live greeting — *"Hey, what's up everybody? We're here live streaming from Panama"* — and is structured as an unscripted two-person working session rather than a produced tutorial. - **Screen-shared workflow walkthrough**: The bulk of the content is a real-time demonstration — creating a Claude skill, letting Claude control the browser to run Midjourney prompts, and iterating live with visible failures ("you can see it's not super smooth… there are kinks"). - **Two named hosts with stated roles**: - Kiran, who introduces himself as *"the GM of Kora"* - Lucas, who introduces himself as *"creative director, just making sure we look as beautiful as possible on the web"* ## Explicit Self-Description / Claims - The session is framed as part of an internal event: *"we're in Think Week"*, with the goal *"to automate ourselves away a little bit"* on recurring tasks. - The channel/organization's stated aspiration: *"moving into more impactful, higher level work and deploying a bunch of agents to iterate on our style and our taste. That's the goal."* ## Demonstrated Audience - **Practitioners of AI-assisted creative and editorial work**: The material assumes familiarity with Midjourney (styles, stylization settings, "vary strong"), Claude skills, browser control, MCPs, "nano banana" editing, and the co-work environment. - **Nontechnical creatives working with technical tools**: Lucas self-describes the editorial team as *"nontechnical people"* who use Claude for visual metaphors when articles are *"dense or super technical."* - **People curious about automating taste-based judgment**: The central recurring question is how to encode subjective design taste — *"how do we help Claude understand a little bit of what we like and what we don't like."* ## Claimed / Observable Scope - The channel demonstrates workflows around an **editorial publication with daily essays**, each requiring a bespoke cover image: *"each day basically our editorial department publishes an essay and we need to give it a cover image."* - The site's content extends back to at least **2017** — *"our oldest things do not have cover images… we don't have the time to go all the way to 2017."* - A distinct **house visual style** is referenced throughout: a Midjourney "profile code," image reference banks, and constraints such as dark website background and 1:1 thumbnail cropping. ## Takeaways the Hosts Explicitly State - *"The lesson is Claude can control your browser and co-work is pretty cool."* - *"If you have something that you do over and over again, you can create a skill… and just start small and build an empire."* - On subjective taste: *"it's really good to just ramble into monologue… so that Claude can make sense out of your ideas."* ## Creator Claim vs. Demonstration Distinction - **Claims**: Think Week purpose, the goal of "deploying agents to iterate on style," and the value of the skill-sharing workflow are stated by the hosts, not independently verified. - **Demonstrated**: The live skill creation, browser automation with mixed results, the iterative critique process (e.g., rejecting "dirty" textures, favoring saturated backgrounds, checking for "extra fingers"), and the real friction of the process are directly observable in the transcript. # Delivery Observation: "Can AI Learn our Designer's Eye?" ## Format & Participants - **Two-speaker conversational stream** between Kiran (GM) and Lucas (creative director), livestreamed from Panama during a "Think Week." - Delivery is **improvised and unscripted**: heavy filler words, self-corrections, and laughter ("Look, mom. No hands."). - The structure is **live problem-solving, not a lecture**: the pair talks through a workflow, builds an automation "skill" in real time, and reacts to results as they appear. ## Tone - **Casual and enthusiastic.** Kiran opens with "Hey, what's up everybody?" and the register stays informal throughout. - Lucas uses **design-critic vocabulary** in a relaxed way: "visually arresting," "AI slop," "pop saturated background colors." - Moments of self-aware humor about job displacement: "this type of stuff is scary because like at the end of the day, like, oh, this is my job." ## Recurring Phrases & Verbal Habits - **"Like" as an extremely frequent discourse marker** — appears in nearly every extended turn (e.g., "so like take a little part of like something we have to do every day"). - **"Kind of like" and "sort of like"** used repeatedly to hedge descriptions: "it's kind of like similar to that in a way," "sort of like get rid of first like all the initial ideas." - **"At the end of the day"** recurs at least four times: "at the end of the day when we're generating images…", "I think at the end of the day uh like the only reason why we use midjourney…" - **"Heavy lifting"** used as a recurring metaphor: "remove the heavy lifting from us," "the heavy lifting again is just the code," "we want the heavy lifting to be done by the the code." - **"You know"** as a frequent turn-ender: "you're making a symphony you know," "it's a skill MD file… you know." ## Teaching Habits & Explanatory Patterns - **Learn-by-doing demonstration** instead of slides: "So we're not creating a skill yet, but we can like experiment because we're here." - **Incremental scaffolding** — breaking the process into named steps: "there's an article going in and we want a prompt going out… that's one step. And then the next step is…" - **Restating the learner's process back to them** to confirm understanding: "Um and then you approve basically and that that's mainly the the whole process. Is that correct?" - **Concrete before abstract**: Lucas grounds abstract taste talk in a specific article example (the orchestra conductor metaphor for the Cursor article). - **Takeaway synthesis at the end** from Kiran: "if you have something that you do over and over again, you can create a skill… and just start small and build an empire." ## Communication Patterns - **Interruptible, dialogic delivery**: frequent overlapping short interjections ("Yeah." "Right." "Okay. Okay."). - **Audience-directed framing**: Kiran narrates visible actions for viewers ("you can see it takes a screenshot on the left"). - **Immediate critique of live outputs**, modeling expert judgment aloud: "I already feel this this one will look maybe weird, but we'll see." - **Explicit anti-obviousness principle** articulated twice: "if the article is about like uh engineering, we shouldn't show an engineer, right?" ## Stable Patterns vs. Isolated Moments - **Stable**: filler-heavy informal register, step-by-step decomposition, "at the end of the day," metaphorical "heavy lifting," live demonstration style — all recur across the full transcript. - **Isolated**: jokes like "Look, mom. No hands." and "I gave my life away" (login moment) occur once each and shouldn't be treated as signature habits. - **Uncertainty note**: laughter is marked twice in the transcript ([laughter]); humor seems present but its frequency can't be firmly established from transcript markers alone. # Position Extraction: "Can AI Learn our Designer's Eye?" ## Speakers - **Kieran** — GM of Kora (identified in transcript) - **Lucas** — Creative Director (identified in transcript) ## Context Live stream from Panama during the team's "Think Week," dedicated to automating recurring work. The session explores building a Claude-based agent workflow to automate cover image generation for daily editorial essays. --- ## Positions and Claims ### The Current Manual Workflow **Lucas** described the existing process: - Editorial publishes a daily essay, each needing a bespoke cover image in the house style. - Workflow: read the Google doc → brainstorm ideas → generate 20–30 Midjourney images → use personal judgment to select and polish. - Images must avoid being "too on the nose" — e.g., an engineering article shouldn't show an engineer. ### Prompting Philosophy **Lucas**: - They built a Midjourney style "profile" from a bank of manually created reference images; the mood board does "the heavy lifting." - Prompts have become simpler — "less than 10 words," just nouns and adjectives; no explicit style descriptors like "Greco-Roman neoclassical." - Words alone are insufficient: > "I don't think words go far enough to explain how visuals could look." - Found images are re-injected into prompts to steer style. - Claude is used ad hoc when uninspired or when articles are dense/technical — e.g., an article on Cursor framed developers as an "orchestra conductor," producing a prompt like "orchestra conducting and opera." ### The Automation Goal **Kieran** framed the two steps: article in → Midjourney prompts out; then prompt → results → feedback → iterate → approve. **Lucas** on the hard part: the self-qualification and self-correction step — > "that's something that I even struggle to figure out how to even talk about." **Lucas** proposed giving Claude the full image bank to "describe in words what makes these images good." **Kieran** suggested starting simpler and building a framework to check whether output images fall within spec. ### Implementation Approach **Kieran**: - Browser control by Claude is "a very good first step" — no API keys needed, users only need to be logged into Midjourney. - Skills can be stored and shared; work in chat and co-work, not code (which is separate). - Co-work can take screenshots, view generated images, and "make arguments why certain things are maybe better than others" — potentially iterating with the "Lucas mindset." ### Evaluation Criteria (Lucas's Stated Taste) - Avoid "dirty" images with heavy textures — less appealing to stare at. - Gravitate toward "pop saturated background colors"; avoid too-black-and-white images that get lost on the dark website. - Check for "extra fingers" to avoid "100% AI slop." - Thumbnail constraint: images crop to 1:1, so subjects must be center-weighted. - Iteration via Midjourney's "vary strong / vary subtle" is valued; ideally Claude would recognize potential and create variants, "picking the best of the best." ### Technical Observations During the Demo - The agent initially failed to append the style code; the pair updated the skill to always append the "profile code" (starting with "B"), prefer fewer words, and stay within the "every vibe." - **Lucas**: Midjourney is "more like a slot machine," not a reasoning model — centering a subject via prompt is unreliable. He proposed adding "nano banana" as a post-editing step (it has a skill and an API). - **Kieran** acknowledged "there are kinks" and "you need to nudge it once in a while"; also noted a possible confusion between the style code and Midjourney's "stylization" setting. ### Tensions and Qualifications - **Lucas** on job anxiety: > "this type of stuff is scary because... if an agent can do it, what's the value of what I do?" — while framing automation as freeing time for higher-impact, big-picture work. - **Lucas** admitted he hadn't actually read the test article and had to re-check it while judging image relevance. - **Kieran** on verification: you know it works by whether the output is something you're proud of, regardless of how it works. **Lucas** agreed he wouldn't care if a different tool than Midjourney were used — Midjourney is used for "context, reliability, and consistency," though he also called it "kind of our identity." - **Lucas** compared the effort to onboarding a new team member: upfront investment despite it being faster manually the first time. ### Closing Recommendations - **Kieran**: start small, capture repetitive work in skills, "build an empire"; for subjective/taste-driven work, ramble into a transcript and let Claude make sense of it. - **Kieran's** next step: transcribe this whole conversation and ask Claude to update the skill. # Speaker Attribution & Provenance Record **Video:** Can AI Learn our Designer's Eye? (ID: 01a0f15a-af1c-7043-81f7-548736369848) ## Speakers & Roles - **Kiran** — **Creator / host**. Self-identified: *"I'm the GM of Kora."* Leads the live stream, drives the automation experiment, and delivers the closing summary. - **Lucas** — **Guest / featured expert**. Self-identified: *"I'm Lucas, creative director."* Explains the editorial cover-image workflow and provides design judgment throughout. No narrator or other identified speakers; the format is a two-person live-stream conversation from Panama during "Think Week." ## Referenced Sources & Tools - **Midjourney** — image generation tool; no official API noted (unofficial/browser-based workaround discussed). - **Claude / "clot" (transcript phonetics)** — used for idea brainstorming, skill creation, and browser control. - **Co-work / Chat / Code** — Claude product surfaces discussed; skill works in chat and co-work, not code. - **Nano Banana** — cited as a possible editing step with an API and existing skill. - **Photoshop** — used for final refinements (logos, text Midjourney "hallucinates"). - **Google Docs** — source of editorial articles. - Internal assets: website mood boards, image reference bank, Midjourney "profile code" (starting with "B"), style settings (16:9, stylize 100). - Test article: *"The quest to be good at everything"* (co-founder of Bark); earlier example: article on the Cursor team. ## Third-Party Claims - Lucas describes their own past workflow and design preferences; claims about Midjourney behavior (e.g., "it's like a slot machine," not a reasoning model) are the speakers' characterizations, not verified facts. ## Promotional Segments - **Implicit product promotion:** the session functions as a demonstration of Claude skills, co-work, and browser control ("the lesson is clot can control your browser and co-work is pretty cool… start small and build an empire"). Midjourney Turbo (paid tier) mentioned. No explicit sponsored segments flagged in the transcript. ## Availability Limits - **Comments:** no comment availability information present in source material. - **Curated metadata:** none supplied beyond title and transcript. - Speaker attribution relies on self-introductions only; some tool names are ambiguous due to transcript phonetics ("cloth/clot" = Claude; "every" possibly a brand/style name). Hello, welcome. \>\> Hey, what's up everybody? \>\> We're here uh live streaming from Panama uh from the beautiful beach and I'm here with Lucas. Uh my name is Kiran. I'm the GM of Kora. \>\> And I'm Lucas, creative director, just making sure we look as beautiful as possible on the web. \>\> And we're in Think Week. And what we're trying to do today is we're trying to automate oursel away a little bit. So like take a little part of like something we have to do every day or every week, see if we can automate it away, make it a little bit easier. And I'm going to gem with Lucas today. So Lucas makes these beautiful images and he has mastered the art of the midjourney styles. \>\> Yeah, you can see them here in the website. This is the the images that you're familiar with already. I guess this type of stuff. Yes. Okay. So we're trying to like like see if we can have like an agentic way of creating some of these like we're not going to automate everything away but like it's still manual because mourney doesn't have an API an official API and we're just seeing like can we have cloth or something uh like cloth code or co-work control the browser or is there some like API uh way like unofficial way to work with images so that your life is easier basically. So can you explain a little bit like what you want to try to achieve? \>\> Yeah. So like each day basically our editorial department uh publishes an essay and we need to give it a cover image each single day and obviously we are not going to like repeat ourselves. We are going to create bespoke images that are relevant and specific to each article. Um, and we want them all to be in like our every style. But that means that we're usually going through the process of reading an entire Google doc, like understanding what the piece should be about, coming up with a few ideas and then experimenting a ton on mid journey until we find something that feels like visually arresting enough and then we can iterate on it and like uh basically polish until we feel like it really captures the spirit of what we're trying to say on that specific article. Um, so my goal at least is that we are able to create a process in which at least uh an agent or a clot or something like that uh is able to read an article, propose ideas uh in a way that's not uh like too on the nose or like uh like if the article is about like uh engineering, we shouldn't show an engineer, right? We want to like go a little bit deeper than that. And uh and besides that just how do we help uh claude understand a little bit of what we like and we what we don't like because it's at the end of the day when we're generating images in mid Journey for each article we're generating probably like 20 to 30 different images and then we're using our own judgment to sort of like select what fits best the article but also what fits best every and then we're like polishing on like what we think. So the what would be ideal honestly it's like how can we get it to understand the process from Google doc to ideas to prompting but also like after prompting like how do you like self-qualify and how do you sort of like selforrect and iterate or like select what's better or what's not and that's something that I even struggle to like figure out how to like even talk about it but yeah \>\> exactly so basically what we have already is like we have a like a a style in midjourney Yeah. \>\> And like normally if you like so how the process is you you go brainstorm a little bit from you have like an article and then you say hey can you explore some ideas from this written text and then do you ask already to cloth or judging to come up with these prompts and then attach this style like what is your workflow currently? So usually um like we've basically built this um profile that's like uh just a lot of references that we used to do manually. So before like our images used to be uh much more manual because we would create like individual assets and then we would like add backdrops and we would like remove backgrounds and we would like create the composition. But we did that enough that we were now able to create like a whole bank of images for mid journey to pull context from. So now we are really trying to like remove the heavy lifting from us and leave it to midjourney which means that our prompts are becoming simpler than ever. So we are for example we're not saying like noise and like Grec Roman classical neocclassical style or anything like that because the model should already sort of like understand that in a way. Uh, so we're like going for very simple nouns, adjectives, and like probably we're trying that our prompts have like less than like 10 words usually. uh which is not like what people would have expected like but we are letting sort of our our mood boards do a lot of the of the heavy lifting and then also our image references because at the end of the day like I think uh \>\> like what the word big for me might feel different to Kieran. So like for example like words I don't think uh go far enough to explain like how visuals could look. So whenever we found like one image, we'll use it again in the in the prompt and inject an image in the prompt to kind of like keep steering the style. Um so we basically sometimes use clot when we are feeling um like uninspired when we read an article and it's very dense or super technical maybe for us nontechnical people and then claude will help come up with like visual metaphors like for example this article is a great example like what the team behind cursor knows about the future of code and basically this whole article goes around uh explaining how like we should think about ourselves as sort of like an orchestra conductor because you are sort of like uh controlling many different aspects and you're not just like playing an instrument but you're making a symphony you know you're sort of like connecting a lot of different parts so relying on these types of visual metaphors goes a really long way so for example for this prompt specifically I believe it was just like uh orchestra conductoring and opera something like that and like we just kind of like went hard on it maybe like changed a few words here and there when we're like experimenting on the on the specific prompt um and yeah that's kind of like how we how we got like specifically that image. \>\> Yeah. So like let's say okay this is our Asian native guide from Dan. It's not an article like should we take an article create the image or can we just do like an example of something? \>\> I think uh let's just grab an a random article from the past. \>\> Okay. Let's let's grab one that doesn't have images like this maybe. \>\> Yeah. Grab uh \>\> like where \>\> or you know what what would be great? Go to like newsletter. So at the very top uh sorry go at the very top like uh \>\> you do yeah I'll go to recent essays you'll notice that uh our oldest uh things do not have cover images because we cannot we don't have the time to go all the way to 2017 and \>\> let's create some for these \>\> so we can like basically find something that we think it's interesting maybe \>\> I don't know like I'll just go to like page four this this one. So what I would do basically would be uh maybe first sort of like try to come up with ideas just at looking at the title and the by line. So the quest to be good at everything and then the co-founder of bark reveals the system he uses to manage two growing businesses and family. So it's sort of about the goal and the the systems that somebody uses to sort of like be very uh successful in a way with their business and stuff like that. So a lot of the different like visual metaphors that you would want to like think about is like maybe somebody working hard and like all the low hanging fruit. So like we start with the low hanging fruit so that we can kind of like get rid of first like all the initial ideas and then try to like get a little bit deeper. Okay. Okay. Yeah. \>\> Okay. Wait, wait, wait. So um this all goes on in your head like do you already like there's no no skill or flow or anything like that currently that does all of that? currently. So \>\> sometimes we would select the entire article and just like the copy copy paste it into cloud and be like give me 10 ideas when we're doing like really bottleneck. \>\> Yeah. Okay. So what I'm what like what I'm uh going to do is I'm going to extract all this information from your brain \>\> and then um I'm going to like dump all of that as a transcript and see if we can like extract something from that. So maybe like okay so we have an article um you read the article this is how you think about it any other pointer so you you read the article like any other ways you interpret this from like thinking about brainstorming about like what like what you're like wanting to put in that prompt because really the the thing is there's an article going in and we want a prompt going out a midjourney prompt And that's one step. And then the next step is like that prompt goes into midjourney and we get results and we then give feedback and iterates. That's the second step. \>\> Yeah. \>\> Um and and then you approve basically and that that's mainly the the whole process. Is that correct? \>\> Yeah. It's uh unless we need to like add something extremely specific in which case we'll have a like a final step of refinement or like overlaying something like some logo that mid Journey will like hallucinate or \>\> or whatever like if we need to add text Journey can't do that. So we'll do some stuff in Photoshop still. Yeah. \>\> Um \>\> yeah, I think like that's pretty much the steps. \>\> Okay. Uh do you want to Yeah, let's let's see. Let's uh Okay. Well, okay. Let's start creating something. So, uh let's let's do a new chat. So, we're in the chat now in cloth. What I want to do is see if we can create a skill with this outline. So, this is basically what we just talked about. Can you analyze this and can you make a workflow of steps and things that need to be done here? So the the the input is an article and the output should be maybe a set of midjourney prompts that we can run. Um and yeah, then we should go from there. Yeah. So what I normally do is like talk through something like this. go in here. Um, you can let it research and pull in other information as well. So, if you say like, hey, we have best practices or a style guide or something, you can just also put that in. \>\> Yeah. \>\> Because why not, right? \>\> Right. Like the prompt should be smaller, whatever. Like \>\> Yeah. Yeah. Exactly. Yeah. Or if you already have something that you sometimes use uh like to brainstorm already like like a command or something like that. And what we're going to do is make an article to midjourney prompts uh workflow first and then the next step is to have it also interact with the browser. \>\> So it controls your browser which is a very easy way in. You can do MCPs and COIs and all of that but you can also say can you just control my browser. It's a very good first step to see if it works and also it's very familiar because you can just see it doing things and you can go in and just take over if you want to, \>\> right? And like while it's using my browser, can I do something else in another tab? \>\> Yeah, you can just let it do its thing. It says this this tab is controlled by clause and it's just doing its thing. Um, so yeah. Okay. So, it's uh doing this. So I think we need to also tell it that this is the style it needs to use. \>\> Yeah. \>\> Okay. \>\> Append it at the end. \>\> So can you append this style also to every prompt at the end? And also the the style needs to be set to 100 stylize. Stylize right here. Yeah. \>\> Settings. \>\> Want to be 16 by9 cuz right now you're 3. \>\> Oh yeah. Okay. \>\> Uh 16 by 9 stization 100. Okay. Okay. Great. Okay. So that's good. Uh and 16 by Okay. So that is then going to be included. Okay. So, so what we can do now, we can actually see if this kind of works. So, we're not creating a skill yet, but we can like experiment because we're here. We have the context loaded. Um, and we can say, oh yeah, let's try it out maybe. Um, or we can already create the skill. Maybe let's create the skill first and then we can iterate on the skill. \>\> Okay. \>\> Okay. Can you create a skill for this? And the next step should be that it uh tests this out and runs these prompts inside midjourney and then ask me for feedback on the generation so we can iterate on it. So we'll create the skill and the skill you can store to your account and then it works in co-work and chat. So we'll we'll probably jump into co-work to try the skill out and we'll have the chat here to work on the skill. So it's like a little bit of a meta work. But in co-work is a little bit more flexible doing these things and it can also pull in the images it generated. So you can also have it look. \>\> Yeah. And it can also look at the images better and you can say hey can you look like can you see if any of these things are true? So it can like already like give a pass on images because it has vision. \>\> Right. Right. Right. \>\> Like it takes screenshots of the screen. It takes screenshots of the screen and can already like ar make arguments why certain things are maybe better than others and you can say hey can you do 10 iterations into this direction. So if in the future we can capture your thinking and reasoning, we can embed that in the skill and say can you iterate 10 times with the Lucas mindset, \>\> right? \>\> And then maybe it's better. \>\> And do you think it would be helpful also maybe for like cowwork to have like uh just like the bank of like what's a good cover image \>\> at least so that like when it's outputting stuff at least it's like oh this is like too far away. This is like way closer to like things that have been done before that feels more aligned and less aligned. \>\> I I think not the full bank probably, but it's we can give it the full bank and say what's like describe in words what makes these images good, \>\> right? and maybe have a framework that you can uh have an input image and then say if it's within this spec \>\> or not like we we but I think let's start simple \>\> let's see maybe it works like like you said now you just have a few words with the style and it's like kind of working \>\> so yeah so let's let's hope that it's working as well so it's creating a midjourney automation skill now \>\> and skills can be shared so you can share this with anyone The only thing they need to do is they need to be logged in to midjourney of their browser. That's literally the only thing they need. There's no API keys or anything. Uh you can say see it here. We could add references. So we could add all these images as reference and then cross check. So there's like it's extendable. But let's start. Um I just hit copy to skills here. And what it does then it's saving to my skills and now it's available inside chat and co-work \>\> and code too or no \>\> no code as a separate piece. \>\> Okay. Okay. \>\> Um but it's available here. So okay let's see. So, we're going to take this article and we say reads and use the let's see how the skill the journey workflow journey flow skill typos and can you generate some cover images for us? Okay. So, I clearly needed to use monologue here. Okay, we're queuing this and see what it's going to do. I \>\> mean, this type of stuff is scary because like at the end of the day, like, oh, this is my job and like if an agent can do it, what's the value of what I do? But it's also sort of like that's also sort of the theme of today's think week, which is like how can we promote ourselves and move away from like \>\> the the dayto-day and focus more on big picture type of stuff so that we can even have more impactful work versus getting lost in the week. Yeah. Right. So like like if you don't have to do these things, you have more time to like generate more of like what our style is like more like you can do other things. But yeah, but I I I think you asked me a good uh question in the beginning. You were like how do I know like this is actually midjourney? How do I know this is correct? Like \>\> right \>\> that's a good question because I I even think if you vibe code or do something like that how do you know like well either you read the code but some people don't code \>\> I think you know by does it work for you like is it is the is the output something that you're proud of that makes you happy I think if it yeah is it something that is just you can put your stamp on where you're like I'm proud of this this is good I think then it works like it doesn't matter how it works. Even though like even though like would you care if we use like a different like what if we have like a a fine tuned version of this style that is not midjourney. Would you care that we use midjourney or not? \>\> Probably not. Probably. I think at the end of the day uh like the only reason why we use midjourney is because of all the context that exists in it and like the reliability and consistency. But if we are able to kind of like recreate the consistency and reliability in some another thing totally. Yeah, exactly. Yeah. So, but but you but you have this thing like I need to use mid journey because that is kind of our identity which I guess \>\> it is. Yeah, exactly. Yeah. Okay. So, it makes a plan now and it's running the things. Okay. So, see my browser goes over. \>\> That wasn't you. \>\> I didn't do anything. I'm not controlling anything. \>\> Oh \>\> Okay. Let me log in first. Okay. We're logging in. I gave my life away. Um, and let's see. Continue. Continue. Okay. So, I'm not clear. Hands here. Hands free. \>\> Look, mom. No hands. \>\> So, okay. We allow it to do stuff like browsing and stuff like that. So, that's safety. Doesn't like do stuff without you wanting to um Let's see. Let's put it a little bit here so we can see both sides. Um, go to mid journey finish. Okay. So, you see like it's not super smooth, \>\> right? There's still kinks. \>\> There are kinks, but like uh Okay, \>\> it does work like like you need to like kind of nudge it once in a while, but also like we do this in a browser. You could totally like create something around the API and stuff like that. So you can see it takes a screenshot on the left \>\> to like figure out where to type. \>\> Yeah. What we do is low. Okay. So you're like, "Okay, it's loaded. Good. I can do stuff." Um I can see the top. Let me enter the first prompt. So and and this might be pretty That's a long run. Ethereal 3D render. I already feel this this one will look maybe weird, but we'll see. \>\> And you can see it's pretty slow now, but we can optimize this. We can like already tell it to do things different. Yeah. \>\> Yeah. And there's also Turbo, which if you pay more for mid Journey, you can \>\> Okay. You can go faster. Yeah. So, like it's funny because you already have feedback, right? \>\> Yeah. From the prompt already. Like look at that. \[laughter\] \>\> And it didn't append. It didn't append mega \>\> Okay. Okay. Okay. Okay. So, we're we're supposed \>\> look like the Thanos stones. \[laughter\] \>\> Uh, okay. So, we we go back to chat for a second here. Go. Uh, yeah. So, here and uh we should always uh append. Okay. So make sure the skill is very particular about always appending the style that we gave and also less words is probably better and also it's for every so whatever the every vibe so don't do non every vibe things and can you update the skill please I don't know if it's going to get confused because you mentioned this use the style to append it but there's also a thing called stylization which is \>\> our thing is technically called profile code. You know, \>\> make sure to add the profile code that we gave \>\> that starts with a B. Yeah. So, we have this here. So, this is Yeah. Okay. The mood \>\> you can just copy paste that. \>\> This is it. \>\> Yeah, that's it. \>\> It's shorter here. \>\> Yeah. Mid is sort of like truncated or it's kind of like a bitly. \>\> Oh, nice. Okay. Okay. So, we'll So, this is like like already the process. Um, we'll do this as well. And it's not loaded in here, but we're just going to actually go co-work again. um about this conversation can continue. \>\> I guess while you do that um it's kind of interesting because this sort of like is a very similar type of process when you're like on boarding a new team member like they'll get it wrong many many times and and like yes like for sure if I did this like I would already be done manually you know like we wouldn't be figuring all of these kinks but it's kind of like you're investing all of these extra time so that like you don't have to like keep doing it forever and ever and ever so it's kind of like similar to that in a way. Yeah, for sure. And and you read this prompt, you're like, this is this is like not correct. But then I ask you why, and you're like, well, like maybe it describes too many things. \>\> The heavy lifting again is just the code, not the not the problem. \>\> Yeah, exactly. Yeah. So, um Okay. So, we want the heavy lifting to be done by the the code here, by the profile. Uh so just like say more like swers we don't care about how stylistically it looks as much right like we don't need like a heavily visual prompt. We just need need uh concepts. Yeah. You know like oh we're talking about whatever this other Yeah. Yeah. Yeah. Yeah. Yeah. For sure. Okay. So, \>\> okay. \>\> It's already going this already. \>\> Leather notebook, morning light, coffee steam, reflection moment. This is already better for sure. \>\> Uh, oh, and it run multiple prompts. \>\> Yeah, it's running all of them now. Yeah. So, you can see already like this is already like probably an improvement if you can just say like, hey, I have an article and then it runs like five things. You come back and you review it, right? and you say, I like this one for these reasons. Can you do another iteration? And then it just does more and you look at it. So you feel already, hey, that's like a different way of relating to this kind of work. Instead of like you need to be active and wired, you can now just kickstart it \>\> and come back more like you you have someone do it for. \>\> Right. Right. And then like how would you think like the feedback process would work with Claude? like does he see my cursor while I'm pointing like this image looks better or like do I have to like attach it like in a conversation to like kind of get \>\> so yeah so it knows the prompts \>\> okay so I can talk about the prompt and like they like the fourth image in that prompt \>\> yeah you can also probably just click on here and copy this image and say I like this one the best \>\> okay \>\> for these and these and these reasons and it can browse to that place and then it knows what you reference \>\> right \>\> so it doesn't know like the the the codes but like if you give the link and say I like this this one. \>\> Yeah. \>\> So how like like looking at this what are you thinking like what what is what is going on in your head? \>\> Can can I scroll or interrupt? \>\> I I mean it will interrupt. \>\> Okay. So, I don't want to scroll yet, but I think like some of them are potentially feeling like a little bit um like not related to the article. Perhaps like when we see like a family or something like that, like maybe it's not super related to the article, but maybe like this idea of a human supported by clockwork system, whatever. Like maybe it's a little bit more interesting or visually arresting that I might be inclined to kind of like keep exploring here. But I also feel like I have to look at the article again cuz I haven't read it. \>\> Yeah. Yeah. Yeah. And also that step we can make way better. Like we can like your thinking of you reading an article like we can capture that and make that prompt way better. \>\> But this is just like a framework \>\> and now it's like making \>\> and it's like okay some things look great. Like for example, this uh top image in the right with a guy holding the thing looks awesome. But the only issue with it is that like our images in our in our website uh when they are on a thumbnail like before entering the actual article they get uh cropped to one to uh one by one aspect ratio. So we really need them in center be center. \>\> Yeah. And we need that in the prompt then. Okay. This one is awesome. Can we iterate on this but make sure that the subject is in the center. So it also works in a one by one ratio. \>\> And it's interesting because I I personally never prompt for things to be in the center because I feel like midjourney is not like nano banana. It's not like a reasoning model. It's not going to be like oh I need to move this to the center or whatever. It's just like less way less deterministic. It's more like a slot machine. But we what we could do is we could add nano banana as another step like an editor \>\> like a Photoshop step like you're like I love this but I want it \>\> subject to center. \>\> Exactly. Yeah, that could be it. Um because like most of the time Oh, and it was the 16 by9 aspiration. \>\> Yeah. Okay. So, it's getting confused but we can like \>\> Yeah, we Yeah, but you can see already like where it's working, where it's not working. What I what I would do is just like copy this skill and start hacking on this on your own computer and seeing how far you can bring this. \>\> Yeah. Like for example, the guy laying on top of like the yellow gears. Like I think that uh composition is like kind of really cool, you know? I I feel like it's something that I could imagine in like a art deco type of sculpture or something like that. And and maybe there's something to that idea of like the article which I kind of like forgot the title. What was the title? uh \>\> uh the quest to be good at everything maybe and it's all about like manage to grow in uh businesses and family and all that sort of \>\> okay there's an article \>\> um and yeah maybe it's all all about the systems that you need to kind of like become better as a person or as a businessman and maybe like this it's kind of like feeling like that you know \>\> um so I could see myself using this image but then if I need it in a 16x9 I would just sort of like drag it here or I ask Nano Banana to like add that step But we could like there is a like nanob banana has ski like there's a skill for nanob banana and it has an API. So it's super easy. You can say okay I want this image but create a one by one and this and this and this and this \>\> after you approve it and then it just exports all of those and that's it. \>\> Yeah that's sick. Yeah. Yeah. That's sick. \>\> Um that's interesting. And then can we go back to see the rest of the outputs? \>\> Yeah. And can we scroll? \>\> Yeah you go. \>\> Yeah. Yeah. Okay. Cool. Because like let's go through the beginning. So okay. So here you can see like how our code is not here at all. So like this is kind of like useless for us in a way. But then like these images Oh, it's still sort of like controlling the browser. It's still doing stuff. Stop. Stop. Okay. It's still working. Yeah. Yeah. Sorry. Sorry. Taking. So you can always say stop here and it will stop stop working. So now you awesome. And you can also open another tab and then just take control and let this tab do its thing as well. So you can just do do that. But keep going. So, for example, like I would say like um and this is all good because hopefully like plot will will learn from more iterations, but like for example, like the dirtier an image looks just because of like heavy textures and things like that, like the the least appealing it is to kind of like stare at it for a long time. So, I would probably like not even look at that. But from like this batch or this initial batch that's kind of like similar, I would like gravitate towards the most like pop saturated background colors and stuff like that. and obviously like double check for like extra fingers or whatever just uh to make sure we're not outputting 100% AI slop. Um so from all of these ones with the hands, I think like the most promising ones are probably this one \>\> and also perhaps this one. But maybe this one is just too black and white and maybe we'll get lost on our website because our website has like a very dark background. So we kind of like want them to stand out and pop. So I think this one could really work honestly. And uh and I sometimes I would just iterate from it because like the journey lets you vary it strong and very subtly which lets you sort of like keep exploring that direction. So I think like ideally eventually we get to a stage where like Claude recognizes that something has potential and then like creates varants of that. So like when we are about to like pick one we're just picking the best one from the best the best of the best basically in a way. \>\> Yeah. Great. Okay. Well this is this is amazing. Um yeah, we're going to explore more and maybe update update us whenever you have this working for you or like whatever you learn from this. But the lesson is cloth can control your browser and co-work is pretty cool. And if you have something that you do over and over again, you can create a skill and capture all that inside a skill and just start small and build an empire. And on top of that, I think like uh with visual stuff sometimes where like it's uh more about taste and subjectivity, it's really good to just like ramble into monologue or just like ramble so that like Claude can make sense out of your ideas because at the end of the day, all of these super subjective things are like a little bit more difficult to articulate and just having something like CL make sense out of that for like your prompting or whatever feels like total like unluck. \>\> Yeah. And really what I'm going to do like whatever like I just transcribe this whole conversation with Lucas and dump it into here and say can you update the skill \>\> right \>\> and and then it's already a better version like \>\> how do how do I access this skill? \>\> Yeah this is locally now on my computer but I will like I can download it here an MD file and that's it. \>\> So it's a skill MD file and there are also some references so you can add \>\> Yeah. So, it's a like a bundle of files, but a main skill file that you can easily read through and see. Um, and I can just send it to you uh as a airdrop or whatever you we want. Like it's it's easy. Um, and yeah, you can build build that out, share it, but it's just a text file. Yeah, \>\> this is really cool. I think it's great. It's a good first step. \>\> It's also really cool to see how you think and go through these things. \>\> Yeah. Yeah. Yeah. It's interesting for sure. So soon the whole website will have new images. \>\> We'll see. We'll see. Hopefully the idea the idea is that we are uh moving into uh more impactful higher level work and uh deploying a bunch of agents to uh iterate on our style and our taste. That's the goal. Yeah. You know. \>\> Great. Okay. Thank you. \>\> Great. Thank you.