The argument in one line.
By connecting Higgsfield's AI media generation to Claude via CLI and building reusable skills, you can automate a complete creative production pipeline that ideates, generates, and tracks hundreds of ad variants weekly without manual work.
Read if. Skip if.
READ IF YOU ARE…
- A solopreneur or small agency owner with basic Claude experience who wants to automate product photography, ad variants, and video generation without hiring a production team.
- A marketer or founder running 2-5 brand launches per year who's comfortable with CLI tools and wants to compress weeks of creative work into hours using AI.
- Someone already using Claude regularly who has a Higgsfield account and wants to see a concrete workflow for scaling from chat-based prompts to scheduled, autonomous content production.
SKIP IF…
- You need photorealistic human talent, complex narrative storytelling, or brand work that requires client feedback loops — this focuses on product-forward, motion-driven outputs.
- You're not comfortable with command line interfaces or don't have time to set up custom Claude connectors and Google Sheet integrations.
- You work in highly regulated industries (pharma, finance, legal) where AI-generated creative carries compliance risk or requires human approval before deployment.
The full version, fast.
Higgsfield's CLI plus Claude turns one operator into a creative agency that researches a brand, builds a product catalog, and ships image and video ads on a schedule. The mechanism is a three-layer stack: connect Higgsfield to Claude via a custom connector for quick one-prompt generation, then graduate to Claude Code with the Higgsfield CLI, the Google Workspace CLI, and a research markdown file that injects copywriting expertise the model lacks. Inside that project, a Google Sheet logs every generation, a creative slate matrix mixes variables across angles and platforms, and reusable skills lock in winning recipes like hypermotion product videos. Scheduled routines then ideate on Sundays and generate overnight, so you wake to hundreds of ready ad variants for testing.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:56
01 · Cold open + promise wall
Promise stated (Claude + Higgsfield = scale-up creative agency) and proven by walling up five-minute outputs: Murmur headphone ads, Sleep Support bottle ads, hyper-motion videos. Sets the bait: 'all of this from one prompt.'
00:56 – 02:47
02 · Connect Higgsfield to Claude.ai (custom connector / MCP)
Walks through claude.ai → Settings → Connectors → Add custom connector. Paste the Higgsfield MCP URL, OAuth in, set permissions. Demonstrates the connector is live and ready to prompt against.
02:47 – 04:50
03 · One-prompt brand generation in Claude.ai
Single prompt: 'Build me a headphone brand from scratch — do research, branding, catalog, and for each product generate a product photo, IG ad, and UGC video via Higgsfield MCP.' Claude returns a brand called Murmur with positioning, target buyer, voice, visual identity, and three SKUs (over-ear Halo, wireless earbuds Drift, open-back wired).
04:50 – 06:20
04 · Review the auto-generated catalog + iterate inline
Walks the Halo / Drift / open-back outputs: product photos, Instagram ads, UGC videos. Shows how to fix mistakes by replying in-thread ('two headers — remove one'), how 'Animate' takes the same prompt graph into a video, and what the video model gets right (realism) vs wrong (duplicated text).
06:20 – 07:20
05 · Marketing Studio: hyper-motion launch video
Asks Claude to use Higgsfield's Marketing Studio to make a launch video for the Halo. First pass is too quiet/intimate; second pass with 'hyper-motion variant, 16x9, more engaging' lands the cinematic ad seen in the intro. Brief detour about a 'sensitive content' block and how to debug the prompt with Claude.
07:20 – 09:15
06 · Second product run: Sleep Support from a reference image
Drops in an existing product photo (blue Sleep Support bottle), asks for Instagram-ready ads. First pass loses the on-bottle text — lesson surfaced: 'be more specific about telling it not to change the reference image.' Iterates to ads with headlines like 'Asleep in 12 minutes' and 'Stop counting sheep — start sleeping through the night.'
09:15 – 10:13
07 · Pivot to Claude Code (desktop) for power features
Why move off claude.ai: more control, reusable skills, true automations. Important nuance — not the CLI terminal, the Claude Code desktop app: still has chat + project, with skills/files/routines layered on top. Sets the scene for the Claude Code build-out.
10:13 – 12:00
08 · Install the Higgsfield CLI + agent skills
Open a blank folder ('Higgsfield Studio'). Grab three commands from Higgsfield > MCP & CLI: install CLI, run hf auth (browser OAuth), install Higgsfield agent skills. Paste all three into Claude Code in one prompt, let it run. Side-explainer: CLI > MCP for token cost and agent efficiency.
12:00 – 13:45
09 · Bring in outside expertise as a research markdown
Pre-built advertising_masterclass.md (617 lines) via a deep-research prompt: best organic ad strategies for 2026 across TikTok / Meta / X, what captures attention, what converts, platform differences. Lives in the project so agents reference it when ideating. Joe-relevant: 'utilize other people's expertise' — swipe-file thinking.
13:45 – 17:55
10 · Master Sheet: log every Higgsfield generation via GWS CLI
Asks Claude to read all 45 assets from the Higgsfield account and write them into a Google Sheet (tabs: generations, by product, by style, planning). Builds a creative-ops database — product, style, image/video, model, prompt, status, result URL, job ID. GWS CLI is the unlock: lets the agent move across Sheets/Docs/Gmail/Calendar/Drive without MCP overhead.
17:55 – 20:35
11 · Test matrix: 30+ variants ideated from masterclass + data
@-tags advertising_masterclass.md and the existing generations, asks Claude to mix-and-match variables (header, style, content type) into testable variants. Sheet gets a new 'creative slate' tab with priority-ranked ideas across products. Frames the philosophy: 'we’re not the bottleneck on creativity or production anymore.'
20:35 – 23:30
12 · Generate rows 3–7 + add status tracking
'Create prompts for rows 3-7, add a status column, generate them in Higgsfield, then mark them complete on the sheet.' Sponsor break for Glido (Nate is joining the Glido team, switched from Whisper). When the batch returns, the Sleep Support bottle drifts off-brand because the reference wasn’t locked.
23:30 – 25:30
13 · Lock the brand asset + regenerate
Drags the canonical Sleep Support product image directly into Claude Code: 'every advertisement must show the product exactly like this — same color, same text, don’t change anything.' Regenerates. New batch comes back consistent and on-brand, mixing nano-banana-2 and gpt-image-2 across angles (curiosity, contrarian, pattern interrupt, stat flash).
25:30 – 28:36
14 · Skills: reverse-engineer a recipe from a winning prompt
Definition: a skill is a recipe for an agent (the pancake analogy). Workflow: find a generation you love, copy its exact prompt back into Claude, say 'turn this prompt into a skill that lives in .claude/skills so anytime I ask for a hyper-motion video you use this.' Highlights the meta-loop: bake learned constraints (e.g. words that get flagged for sensitive content) into the skill so the agent improves over time.
28:36 – 31:00
15 · Invoke the new skill on the locked reference
@-tags the saved Sleep Support reference plus tries /hypermotion. First run uses the wrong skill (default higgsfield_generate). Restarts the Claude Code session; new run reads the .claude/skills/HyperMotion-Video file, asks the right clarifying question ('model in the ad, UGC, or product only?'), and produces the cinematic hyper-motion clip he wanted.
31:00 – 33:20
16 · Output review + honest critique
Output is strong — cinematic feel, real-looking product — but the label text is mangled in image-to-video. Nate names the limitation directly: 'this is the worst AI video generation will ever be,' suggests a workaround (simpler label for hero shots), and pushes back on knee-jerk 'AI slop' framing.
33:20 – 35:27
17 · Routines: schedule the agency to run while you sleep
Claude Code Routines inject prompts on a cadence. The proposed pipeline: Sunday night routine — analyze the Sheet + platform data, add 50 new generation ideas. Monday morning routine — pick 30 blank-status rows, generate, mark complete. Scale to twice-weekly planning + generation. Optional final step: pipe winners into Potato or Meta Ads Manager for full auto-posting. CTA: like + watch the routines deep-dive.
Lines worth screenshotting.
- Connecting Higgsfield to Claude via MCP turns a brand brief like 'build me a headphone brand' into a full product catalog with photos, Instagram ads, and UGC videos from one prompt.
- The difference between using Higgsfield manually and routing it through Claude is the difference between one-off content and a consistent, repeatable, automated creative pipeline.
- Claude Code with a Google Sheets ad-test matrix and a Higgsfield skill can ideate and generate ad variants while you sleep — turning what used to be a creative agency's week into an overnight routine.
- Iterating on a generated image inside Claude by saying 'remove the duplicate header' is faster than reopening the design tool because the agent already knows the exact reference image.
- Higgsfield's Marketing Studio hypermotion format produces AI video ads that would have required a studio, a paid actress, and days of editing to achieve with traditional production.
- Adding the Higgsfield CLI to a Claude Code project means every future agent in that project inherits the ability to generate images and videos without additional configuration.
- The one-prompt creative agency pattern — research brand, build catalog, generate all assets — is a replicable template that scales to any product category without changing the underlying workflow.
One prompt, a CLI, and a scheduled creative agency
What it teaches
Connecting an image-and-video generation tool to an AI coding environment via a command-line interface — rather than a browser MCP — unlocks reusable skills, a tracking spreadsheet, and overnight generation routines that run without you.
01Cold open + promise wall
- A single prompt can trigger a full brand build — research, branding, catalog, and visual assets — when the AI tool has an image-generation connector attached.
02Connect Higgsfield to Claude.ai (custom connector / MCP)
- Adding a custom connector through the AI tool's settings and completing an OAuth flow is the entire setup required to enable image and video generation from chat.
- Connector permissions can be scoped after setup — either allow all actions or restrict to specific operations depending on how much trust you want to extend.
03One-prompt brand generation in Claude.ai
- A single high-level prompt — 'build me a brand, generate the catalog, produce assets for each product' — can delegate an entire creative production run when the tool has the right connectors.
- The model will make creative decisions (naming, positioning, target buyer, SKUs) from minimal input — the value is in the iteration, not the initial specification.
04Review the auto-generated catalog + iterate inline
- Iterating on generated assets is done inline by replying to the same thread, not by starting over — fixing a duplicated header or changing an aspect ratio is a one-sentence follow-up.
- The Animate button in the tool takes a static image prompt graph and extends it into a video using the same prompt — no separate video prompt is needed for a basic animation.
05Marketing Studio: hyper-motion launch video
- When a generation is flagged for sensitive content, asking the agent to surface the exact prompt, identify the problematic phrases, and regenerate without them is faster than guessing what triggered the block.
- Emotional, high-level terms like 'hyper-motion,' 'engaging,' and 'fast-paced' translate effectively into technical prompts when the underlying model has been trained on those aesthetic patterns.
06Second product run: Sleep Support from a reference image
- Being explicit about preserving a reference image is required; without that instruction, the model will interpret the image loosely and alter product labels, colors, and text.
07Pivot to Claude Code (desktop) for power features
- Moving from a browser chat to a desktop coding environment adds project files, reusable skills, and scheduled routines — the chat interface is still present, but the agent gains persistent memory and automation support.
08Install the Higgsfield CLI + agent skills
- A browser-based MCP connector is quick to set up but costs more tokens and is less efficient for agents; a CLI integration is faster, cheaper, and better suited for automated workflows.
- Installing the CLI, authenticating, and loading agent skills can all be done in a single prompt by pasting the three required commands and letting the agent execute them.
09Bring in outside expertise as a research markdown
- A subject-matter research document pulled from deep AI research and stored in the project gives every agent access to advertising best practices without requiring the agent to rediscover them from scratch.
- Using proven external expertise — Twitter threads, YouTube videos, research tools — embedded as a project file is the fastest way to raise the floor on AI-generated output quality.
10Master Sheet: log every Higgsfield generation via GWS CLI
- Logging all generated assets into a structured spreadsheet — product, style, model, prompt, status, result URL — creates a creative-ops database that agents can query and build on.
- A spreadsheet CLI that gives agents read-write access to Sheets, Docs, Gmail, and Drive is more token-efficient than a stack of separate API integrations.
11Test matrix: 30+ variants ideated from masterclass + data
- A test matrix of 30 or more variants, generated from mixed variables (headline, style, content type, avatar), means the bottleneck on both creativity and production has been removed.
- Referencing both the research document and the existing generation history in the same prompt lets the agent mix data-backed best practices with what you have already tested.
12Generate rows 3–7 + add status tracking
- A status column on the spreadsheet prevents duplicate generation by letting the agent filter for blank rows before starting a new batch.
- Without a locked reference image, the agent will use whatever description it infers from context — which drifts from the actual product with each generation.
13Lock the brand asset + regenerate
- Locking a canonical product image as a project asset and referencing it explicitly in every prompt is what keeps brand consistency across a large batch of generated outputs.
- Different model choices (nano-banana-2 vs. gpt-image-2) within the same batch produce different aesthetic results — letting the agent vary the model is a way to test outputs without additional prompting.
14Skills: reverse-engineer a recipe from a winning prompt
- A skill is a reusable recipe for an agent — built by reverse-engineering a prompt that produced a winning output, then saving it so future runs always start from that baseline.
- Skills improve over time by feeding the agent feedback after each run: 'I liked A and C, I did not like B — update the skill accordingly.'
15Invoke the new skill on the locked reference
- Sessions must sometimes be restarted after a new skill is created before the agent will recognize and invoke it correctly.
- A well-written skill asks a clarifying question before generating — 'product only, UGC, or model in the ad?' — rather than making that decision silently.
16Output review + honest critique
- Current image-to-video models reliably produce cinematic feel and realistic motion, but consistently mangle text overlays — simplifying the label on hero shots is a practical workaround until models improve.
17Routines: schedule the agency to run while you sleep
- A Sunday planning routine adds 50 new ideated rows to the spreadsheet; a Monday generation routine picks 30 blank-status rows and produces them — together they create a self-refilling creative pipeline.
- Connecting the generation pipeline to an ad-buying platform as a final step enables fully autonomous posting once you trust the outputs enough to let them run without review.
Terms worth knowing.
- Higgsfield
- An AI-powered platform for generating and animating product images and videos — offering multiple generation styles including HyperMotion, UGC, and cinematic formats — accessible via a web interface, MCP connector, or command-line interface (CLI).
- MCP connector
- A standardized integration that allows an AI assistant like Claude to communicate directly with an external application — enabling the AI to trigger actions, retrieve data, and control the app as part of a conversation or automated workflow.
- CLI (command-line interface)
- A text-based tool installed on a computer that allows software to be controlled by typing commands into a terminal — preferred over MCP connectors for AI agents because it is faster, uses fewer tokens, and is more efficient for automation.
- HyperMotion (Higgsfield)
- A Higgsfield generation style that produces fast-paced, high-energy product videos with rapid cuts, slow-motion inserts, and cinematic close-ups — designed to capture attention quickly in social media ad placements.
- marketing studio (Higgsfield)
- A Higgsfield feature that accepts a product image or URL and converts it into a formatted ad video — offering styles like HyperMotion, UGC, and unboxing — without requiring the user to assemble individual shots or transitions manually.
- GWS CLI
- A command-line interface for Google Workspace that lets AI agents read and write Google Sheets, Docs, Gmail, Drive, and Calendar — enabling automated workflows that create, update, and query Google documents without browser-based interaction.
- ad test matrix
- A structured grid of advertisement variants — each combining different value propositions, headlines, audience angles, visual styles, and formats — designed to systematically test which combinations produce the highest conversion rates.
- skill (Claude Code context)
- A reusable, project-specific markdown file that gives an AI agent a precise recipe for completing a recurring task — ensuring consistent outputs by defining what to do, what to ask, what models to use, and what rules to follow every time the task is invoked.
- routine (Claude Code context)
- A scheduled, time-triggered prompt that automatically starts a Claude Code agent workflow at a set cadence — used to automate repeating tasks like weekly creative generation or data analysis without manual intervention.
- reference image
- A source photo or graphic provided to an AI image or video generation model to anchor the visual output — instructing the model to preserve specific product colors, shapes, text, and branding elements rather than generating them from scratch.
- UGC (user-generated content) ad style
- An advertisement format designed to look like organic content created by a real user rather than a professional studio — typically featuring a person casually reviewing or using a product, intended to feel authentic and reduce viewer skepticism.
- Gliddo (Glido)
- An AI-powered voice dictation tool that converts speech to text in real time — positioned as a faster, more private alternative to Whisper-based tools, used here for hands-free narration during screen recordings.
Things they pointed at.
05:57toolHiggsfield Marketing Studio (hyper-motion / UGC / unboxing templates)
10:15toolHiggsfield CLI + Agent Skills (for Claude Code)
15:25toolGoogle Workspace CLI (GWS CLI) — Sheets/Docs/Gmail/Calendar/Drive for agents
12:18tooladvertising_masterclass.md — project-local deep-research swipe file
20:33productGlido (voice dictation, sponsor)
23:50tool.claude/skills folder — project-scoped Claude Code skill recipes
33:20toolClaude Code Routines (scheduled prompts)
34:55toolPotato — social scheduling / posting
34:57toolMeta Ads Manager — final delivery destination
Lines you could clip.
00:12
“We can ideate and we can generate a hundred times faster than the average human could.”
23:31
“We're pulling the lever on the slot machine, which is AI. If we don't have guidelines, if we don't have recipes — skills — then they're not gonna be super consistent.”
23:51
“A skill is essentially a recipe for an AI agent.”
32:05
“This is the worst that AI video generation models will ever be. Every day, every month, they're going to get better.”
20:32
“I could set an agent off to generate all this stuff, and then I could go to bed, and I could wake up with a 100 different ad copies and creatives ready to go.”
19:46
“We're not the bottleneck on creativity, and we're also not the bottleneck on production.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogystory
00:00So Higgs Field has access to all of the best AI image and video generation models. And Claude lets us and build custom skills and agents and schedule all of these automations to . And when we combine these tools, we're able to actually scale up our content because we can ideate and we can generate a 100 times faster than the average human could.
00:19So in today's video, I'm gonna show you guys how we're able to turn Claude into a creative agency. So real quick, let me show you guys a few examples that I was able to generate in literally five minutes with one prompt.
00:56I mean, those are incredible. I think that this second one's probably my favorite. Like, the zoom in on the product, the detail, all of the animations in the background, the music.
01:06I mean, this is incredible. Even this detail here. Think about how long this would have taken you if you either wanted to edit this by hand or, you know, shoot this with a studio and with a paid actress.
01:16It would have taken so much more time and resources. And like I said, I was able to generate all of those outputs just by talking to Claude with a prompt that looked something like this. So before I show you guys these conversations and exactly how I did it, let me show you how we connect Hicksfield to Claude.
01:30So if you don't have a Hicksfield account, go to hicksfield.ai and you can sign up. You will have to get on some sort of subscription, but once you're there, you basically wanna come to this page that says MCP and CLI.
01:40And we're gonna first of all connect this to Claude in the web. This is just your typical Claude chat that you've probably been using for months now. You're gonna go to the settings and you're going to click on connectors and you're gonna have to add a custom connector.
01:52So down here you can see that I have Higgs field as a custom connector. So go ahead and click add custom connector, and you're basically just gonna call this, you know, Higgs Field or if you wanna call it something else, I don't really know why you would. And then you're going to copy this command right here.
02:04And back in Claude, you're just gonna paste that in, hit add. Mine says it's not gonna work because I've already done that, but yours will connect. And then basically it'll prompt you to sign in.
02:12So you'll hit configure. It'll take you to the Hicksfield OAuth. You'll sign in with the account that you just created, and then you will now have Hicksfield ready to go.
02:20And then if you wanna click on configure, you can basically change the permissions. So if you only want it to be able to do certain things in your Hicksfield account, you can limit it to do that, or you could come in here and you could just say, hey. Always allow all of these.
02:30I don't really care. So however you wanna get that set up. And now whenever you're in a cloud chat, if you go over here and you look at your connectors, you can see that Higgs field is right there and you can just prompt it to create you images with that or videos with that or whatever you wanna do.
02:43Now it will be able to actually talk to Higgs Field as you can see. So just to start off and show you guys a few quick examples, what I said here is build me a headphone brand from scratch.
02:51I want you to do research, build the branding, build the product catalog, and for each of them, want you to generate assets. So a product photo, an Instagram ad, and a UGC video, and I told it to use the Higgs Field MCP for all of these generations. So yes, what we could have done is done all this research on our own or taken all this research from Claude, put it in a Google Doc, and then went over to Higgs field on our own and found all these different AI image and video models and just generated all of this stuff in the interface.
03:16We could have done that. But in today's video, I'm showing you guys how you can essentially just treat Claude code or Claude as the interface to do all this stuff and in a more consistent and repeatable and automated fashion that's really gonna help you scale much more than if you were to do this all manually. So as you can see, it does the research, it looked at the market, and now it's helped me build a brand called Murmur.
03:36And we have stuff like positioning, target buyer, voice, visual identity, and all this stuff. And it created three different products.
03:42So an over the ear, which is the most expensive. We've got wireless earbuds, and we have open back wired headphones once again. And then it just goes ahead and generates all the stuff.
03:51So all these images and videos that you're about to see were just minimal input. It was literally just, hey. Build me a brand and build these things.
03:59And now we have this first photo of our halo product, which looks very nice. We've got this Instagram ad, which came out a little bit duplicated with the text for some reason. But if we wanted to fix that, all we'd to do is say, okay, edit.
04:12And then we could give it a prompt because it knows the exact reference image and it knows this, what we're talking about. And I could just say, hey, you know, we have two different headers in here, remove one of them. So you could iterate really quick on that there.
04:22And you could also click animate. So it's basically the same type of prompt and we could say, hey, turn this into a video where, you know, the headphones are spinning in a circle and floating in the air, whatever we wanna do. So that's the ad, and then we have this video which is basically just a person listening to music, wearing a hood, and smiling at the camera.
04:40So you can see though that video looks really good. Like, person looks very real. So we've got our second one.
04:46We've got the picture right here. We've got the ad right here. That one looks much better actually.
04:50And then similarly, we have this video with music of a person putting one of them in their ear and then smiling at the camera. And then we have the exact same thing for our final product. We have an image.
05:00We have an ad, and then we have a video of someone using the product as you can see. Then I said take the halo, which was the first one, and I said use Higgs Field's marketing studio to create a launch video.
05:12So Higgs Field in here has a really cool thing called marketing studio where you drop in a product or a link to a product and you can even put in your own custom avatar, and it basically turns it into a different type of format. So hypermotion, sort of like this.
05:26It could be an unboxing. You know, there's other styles. There's UGC.
05:29There's so many different things you can do in here for marketing. Like these hypermotions are super super cool. I mean, look at that.
05:35So I told it to use this marketing studio format inside of Claude right here. So it looks at the marketing studio, it found out what to do, and it comes back with this.
05:47As you can see, it's cool, but it's not like the hyper motion style. It's very quiet. We have that weird scene where the person kind of intimately touches the headset.
05:55Don't know what that's about. But it's it's fine. Right?
05:58So I said, okay. That's a little bit, you know, not what we want. I want you to use the hyper motion variant and I want you to make it more engaging and I also want a 16 by nine version.
06:07So I wanted it to capture more attention and then that's when it comes back and it gives us this one which I absolutely love. I think this one is really good.
06:17Okay. But then for the 16 by nine version, it said that it was sensitive content and it refunded my credits. So I said, okay.
06:24Try again. It did the exact same thing. And then eventually I was able to say, Why did that get denied?
06:29Show me the prompt. Figure out why that happened. It read me the prompt, and then it said okay, think it's because of these words, know, all this kind of stuff, so I'm gonna get rid of that.
06:37And it does it again and then I was able to get this version which you guys also saw in the intro of this video. So it's not perfect, but what we're able to do is we're able to build skills and stuff around this so that this doesn't happen again where we're getting sensitive content blocks. Okay.
06:53So another quick example, I drop in an image. So let's say we already have a product that we wanna start with. I want some advertisements and stuff for that.
07:01So I ask it to, you know, make me a couple of Instagram ready ads for this product, and if it has any questions to ask, so it asked me one thing and I gave it a quick answer, and then it starts generating some stuff. So here's one picture. Now what you'll notice is it got rid of some of the words.
07:14So on the original photo, we had like some extra little captions and subtext right here, and it got rid of that. So we have to be more specific about telling it to not change the reference image at all.
07:26That's very important. But anyways, that's a nice picture. This one's also cool.
07:29It's a little bit more of like a effects type of picture. And then we also have this one which is kind of like an Instagram story, which isn't exactly what I was looking for, but it does look good as far as like the interface and the realism. So it said, hey.
07:40We have a calm one. We have a cinematic one, and we have like a real relatable one. Now we're able to keep iterating with what we want, but I said that's not good enough.
07:47I need these to be actual ready to go advertisements across different socials, and then it comes back. It generates some more for us. So this one has text.
07:55It says stop counting sheep. Start sleeping through the night. It has sort of that cinematic feel.
07:59That one I thought was pretty cool. We also have this one, which I loved. Asleep in ten minutes, fall asleep faster, stay asleep longer, wake up refreshed.
08:06I absolutely love this one as well. And you can see with very minimal prompting. All I said was, hey.
08:10Create me Instagram ready ads. It understands, like, headlines and spacing, and I just thought that these looked really, really good.
08:17And then we had one more down here as well. So the main problem is that it didn't have the bottle appearing exactly as we wanted it to in the reference image, which is a pretty easy fix. Here's another one.
08:27You guys didn't see this yet because it's a lot slower paced and it's not the exact, like, HyperMotion style, but it's still a nice little animation. And then we also have this one which is just, you know, the person kind of looking at it.
08:38So these are decent videos, but they're not ready to go ads. We So had to be a little bit more specific again. So I said give me one that's fast paced and energetic.
08:45It has camera cuts, it has slow motion, it has close ups and then we got the one which you guys saw earlier which was right down here. It took me a few more tries and then we got this one that you guys saw in the demo with the capsules.
08:57I thought this one was really, really good. So what's cool about this is you're able to take a super vague high level idea. You can say words like engaging, cool, fast paced, things that are emotional.
09:08And Claude does the hard work of figuring out the prompting and then sends that over to Hicksfield Marketing Studio which has a nice, like, pre trained model in the back end. And then you get stuff like this in minutes. And then what you're able to do is you can say, hey.
09:20This one was a winning ad. You know, this format, this style, this colorway, I want a bunch of different versions of this to test out. And it will just go through and it will plan out and strategize on the different versions, the different headers, and then we'll just go make all of them.
09:32And now you have so many more pieces to to test, and ultimately once you find your winning combo, you just chuck more budget at that one and really try to scale your product. Okay. So that's how you do it in Claude.
09:42You use the Higgs build MCP. It's super easy. It's great to get some POC done in there.
09:46Now there's a lot of things that we wanna improve. The image stuff. We want more control.
09:50We wanna build some reusable skills. We wanna build some automations. And that's where we're gonna head from Claude into Claude code.
09:56Now don't get scared. We're gonna be using the desktop app just for today's tutorial, and all this looks like is Claude. You have a chat, and you have a project, and you're able to build way more power in this interface.
10:08So I'm gonna show you guys how. So the first thing you're gonna wanna do is you're gonna wanna open up a blank folder. You're gonna go to your finder wherever you wanna have it.
10:16So wherever you wanna have this, you could just have it right in your desktop. You're gonna create a new folder and call that like Higgs Field Marketing Studio or whatever you wanna call it. So in my case, this is called Higgs Field Studio.
10:25So the very first thing that we need to do is get everything set up. So let me show you guys the way that I would prompt this to get set up. I would go into Higgs field.
10:33I'm gonna click on MCP and CLI. And when we're specifically doing this for Claude or if we wanna do this with OpenClaw or Hermes or whatever, we wanna do this with the CLI instead of the MCP. Now the reason being, ultimately, functionally, they can do pretty much the same things, but the MCP has all those tools.
10:50So , it's actually more expensive to use an MCP, and the CLI is just better for agents. It's gonna be faster.
10:58It's gonna be more efficient. Like I said, we're gonna use the CLI. So all you have do is copy these three commands exactly.
11:04I'm gonna copy this. I'm gonna copy this, and I'm gonna copy this. And I'm gonna go into Claude code and say, this project is basically being set up to use Higgs Field, and it's gonna be set up for kind of a creative studio, a marketing studio.
11:18I need you to install the Higgs Field CLI. I need you to then run the auth for me to sign in, and then I need you to install the Higgs Field agent skills.
11:26So here below are the three commands. There you go. I paste in all three commands, and then I just go ahead and run that, and it's gonna get you set up.
11:33It's going to install the CLI, and then it's going to do the auth flow. So it'll open up a tab, and then you will just sign in wherever you created your Higgs Field account, and then it will add the agent skills for you to use. Now here for me specifically, when it tries to install this, it's gonna say, hey.
11:46This already exists, but yours will actually just go ahead and get set up. So here you go. It says the CLI is working now running off login, which opens up a browser.
11:53And as you can see right here, it's basically just asking me if I'm okay to connect Claude to Hicksfield, and you're gonna go ahead and hit connect and then sign in. And now you have been authorized. And back in Claude, it should say, okay.
12:03Cool. You're now connected in. I can see your account, and now I'm installing the agent skills.
12:07So exactly like you see right here. Now while this is getting set up, there is something I wanna talk to you guys about, which is the fact that this stuff isn't magic. This stuff just lets you automate things and ideate.
12:17So my point being, if you're not a master copywriter or advertiser, it might be really tough for you to build amazing tier one advertisement copy and creatives.
12:29And that's why what you can do with Claude is you can utilize other people's expertise and you can bring that in to make Claude code the subject matter expert here. So it's not perfect. You know, a a master copywriter is gonna build a better newsletter automation than I would.
12:41So what you can do is in your project, you can do some research. So here's a chat that I did earlier right before this, and I basically said, hey. I need you to do a deep research on the best strategies for advertising in 2026 when it comes to organic advertisements on platforms like TikTok, Meta, or X, and what captures people's attention, what converts, and how it differs per platform.
13:01So I basically wanted it to create a full markdown file called advertising masterclass that would live in this project, and then I could have my agents look at that when they need help ideating or when they need help analyzing what's going on with our data, and it's gonna help them give us better copy and give us better prompts to feed into Hicks Field.
13:19And this is something that I do all the time. Whenever I'm building different agents or automations, I always leverage Twitter threads, YouTube videos, perplexity research. I utilize information that's out there and proven, and I bring that into my systems.
13:32So basically, this did all that research. It asked me questions, and then it gave me a full markdown file, which now lives in this project. As you can see, advertisingmasterclass.md.
13:40It's 617 lines. I could go ahead and open it up right here, and we could read this, which is a master playbook for organic content. It's last updated May 2026.
13:50We have a cheat sheet. We have different platforms, and we have a bunch of information about how attention is captured. And now all of our advertisements are going to hopefully be better because we've done a research doc like this.
14:01But anyways, now we have the CLI installed. We have been authenticated in, and we have our skills installed. So we're ready to get started trying to build some automations here.
14:10Okay. So I cleared out that chat, and here's the first thing that I would recommend you do. So in my Higgs field, if you guys remember, if I go to my assets, we already have a ton of assets generated.
14:20We have 45 different assets created. We have a few different products. We have a sleep pill thing.
14:25We have a few different headphones that were generated, a few different UGC ads. So here's what I wanna do. Go ahead and take a look at all the assets that we've generated in Higgs Field.
14:35I need help basically creating a log tracker of every generation that we do together just for, you know, visibility into our prompts and our data and statistics and things like that.
14:48And I want you to formulate this into a Google Sheet, a few tabs if you need, but use the GWS CLI in order to create this Google Sheet and organize it based on, you know, like, the product and the, you know, the prompts, maybe however you think that this makes sense. Ultimately, this project is being set up to become a master creative agency, so we need to have a database where everything lives so that we can track stuff over time as we scale up how many pieces of creative media we have.
15:16So the reason why I want to do this is because I ultimately want to get to a place where we have something like this as an example. So I have all my generations here. We can see the product, we can see the style, we can see the image or video, the model, and we can see what we're actually generating.
15:30We can also look at the results, and we can look at the actual prompts for all of them, which is really important. From there, we can analyze which ones of these do we like the best, which ones actually converted the best, which ones, you know, had the best budget or had the best spend on our meta account or whatever it is?
15:45And we can have all these other statistics like by products, by style. We can have some planning. And then based on all this data, and especially if you bring in actual real data from your Google Ads or your meta ads or even just your TikTok or Instagram account.
15:58Claude Code can look at it. It can use the subject matter expertise that you had at research, and then it can plan things. So now it's planning out different versions of ads based on certain things.
16:09So we have different value props, different headlines, different avatars, different styles. And it gives us this test matrix where we now have a 100 different things to test, and they all switch up little different variables, and we can ultimately test way more things now because we're not the bottleneck on creativity, and we're also not the bottleneck on production.
16:27Because I could set an agent off to generate all this stuff, and then I could go to bed, and I could wake up with a 100 different ad copies and creatives ready to go. And And then you could also have an agent that every single week generates a 100 more. So you just have like this unlimited bank of things to test, and it's all based on data, not just random stuff.
16:43So it was looking at the master sheet which already exists, which is this one right here. So I'm just basically gonna go ahead and say, hey. I actually just wanna create a completely new one.
16:52That one was good, but I'm just doing a demo, so I want you to just do it again so I can show my audience how good you are at things like that. And if you guys haven't used the GWS CLI before, it's amazing.
17:05It's another CLI just like this Higgs field CLI is a CLI, and our agents are now able to super quickly look at Google Sheets, Google Docs, Gmail, Calendar, Drive. It can look everywhere, and it's much, much more efficient than using, like, a bunch of MCP servers or a bunch of API calls.
17:22So if you haven't tested out the GW CLI before, it's a huge unlock. I'll drop a full video right up here where I talk about it more. Okay.
17:30So now it has pulled all 45 of those generations as you can see. Let me just open up this sheet so we can take a look at it. We've got all the generations right here.
17:37I'm gonna go ahead and make these smaller. Okay. So we've got these 45 generations.
17:41We can also see by product. We can see by style and planning. So really what I wanted to do there is just show you guys that Cloud Code can now look inside of Higgs Field, see everything that we've done, and not only just see it, but it can pull information like the job ID, the status, the prompt, you know, the sizing, all this kind of stuff.
17:57That's really important. I mean, think about how long this would have taken you to manually go look at all your stuff and get it into some sort of internal database. And then, of course, this is where it's able to build on top of it, and it's able to give us tons and tons of new ideas to actually go and generate more creatives from.
18:11So let's actually go ahead and do that. I'm gonna open back up Cloud Code, and real quick, I'm gonna do an at sign, which is gonna let us tag certain things, and I'm going to tag the advertising master class dot m d file, which is, you know, the full breakdown. And I'm gonna say, alright.
18:27So I want you to look at all of the different generations that we've done. I also want you to read that advertising masterclass doc, and I want you to help me figure out a bunch of different variations that we could create.
18:38You know, we've done some things with different headers, with different styles, with different types of content, and I want to mix and match a bunch of different variables so we can ultimately test a bunch of these and put a bunch of budget behind this and see which one of them spends the best. So use your creativity here, use your best practices, and help me get a bunch of different ideas for, you know, more creatives.
19:01And once we shoot it off with that plan, I probably should have said to put that into the GW CLI, into that Google Sheet. Hopefully, it understands that.
19:08If it doesn't seem like it's understanding that, I'll go ahead and stop it and say, by the way, put that in the Google Sheet, but that is what we're looking for here as a deliverable. Okay. So it looks like this is all finished up.
19:17I did have to end up telling it to put it in the Google Sheet. Wasn't that smart yet. But let me go over to the sheet and we'll see we have this new tab called creative slate.
19:25And here, this one looks better than the previous one. It didn't give us a 100.
19:28We could obviously say, hey, Give us a 100, but this one ended up giving us somewhere in the thirties. So it also kind of showed us priorities. So these obviously need to go ASAP, and then we've got other ones as well for different products.
19:40You can see the Murmur Halo. We've got the sleep supplements. So anyways, but now we have this database that we can look at.
19:44We could also have some sort of status so that when we are having it automatically generate these, it can mark them off as processed or done or whatever it is. And then it will create the prompts and talk to Higgs Field to create them for us. So it's gonna be very, cool.
19:58Sorry about the lighting. I just saw, like, apparently a tornado came through and it got very dark. So I turned on my little light.
20:04Okay. So now that we have all of these examples set up, let's say that we wanna generate these top five priority ones. So rows three through seven, we're gonna generate.
20:12So what I want you to do is now create the prompts for rows three through seven, so those first five that you created. But what we also need to do, which is really important, is we need tracking. I noticed on the sheet, you didn't have any status columns.
20:25So add a status column, and what I want you to do is create the prompts for those five, go off to Higgs Field and generate them, and then mark them off on the sheet. Once they come back and they're done as complete or in review or whatever you wanna do there, Just mark them off so we can keep track of what's going on. So just shut that off.
20:43If you guys are curious about how I'm talking with my voice, then check out the tool in the description. It's called Glido. I officially have joined the Glido team, and I'm super, excited about it.
20:50I've switched over from Whisper to Glido. I just truly believe in the vision that we're building over here. So if you guys wanna support and if you need a voice tool, check out Glido.
20:58It's faster, it's private, and it is gonna be way more authentic, so join the movement. But, anyways, this is gonna shoot off to Higgs Field, and I'll check-in with you guys when we get those advertisements back to take a look at. Okay.
21:07So those five are finished. I'm gonna go into the actual Google Sheets so we can take a look. We can see over here that it's added a status column.
21:13It's marked these as complete. We have the result URL. We have the job ID.
21:16So let's take a look at just a few of these real quick. Okay. So this one is interesting.
21:20It's sort of like a meme. You know, it's eighteen months, no sleep. So there's that.
21:24You can also notice that the image doesn't really look like our reference image at all. If you guys remember, if I come over here, it was this one. No.
21:31Not this one. Sorry. It was this one.
21:34So that's honestly my fault. I didn't prompt it to do so. So we're gonna have to redo another round, but let's real quick just see what else we've got here because what you'll notice is these probably aren't gonna be super consistent, and that's because we're just kind of blindly prompting.
21:46We're , which is AI. And if we don't have guidelines, if we don't have recipes around it, skills, then they're not gonna be super consistent.
21:56So this one, I don't know what that is. This one, is a video.
22:07Very generic. Okay. So what happened is in the prompt, I'm assuming, it basically said, hey.
22:13It is a blue bottle. It says sleep support on it, and that's it. So it's creating these just random looking sleep support looking bottles.
22:19So what I'm gonna do is I'm gonna have to go back into Cloud Code. I'm going to drag in the actual image. This is our actual product image.
22:28This is an asset that we should always be using. So when you're creating these advertisements for the sleep supplement product, it has to appear as shown in this reference image every single time.
22:38It must appear exactly like this. Same color, same text. Don't change anything.
22:43And I need you to go ahead and regenerate those five examples. Also remember, the goal here is conversion. The goal is to get someone to want to buy our product.
22:54So go ahead and do those five again. Now, obviously, we're looking for some sort of consistency, but what's important is we have a different kind of angle.
23:02You know, this one was curiosity. This one was contrarian. This one was a pattern interrupt, a question, and a stat flash.
23:07So that's the value of being able to generate so many different types of angles. But also what you'll notice here is it used different models.
23:15It wanted to try these two with nano banana two. It wanted to try these two with gbt image two. So we just have a lot more to play with here, and it's obviously way more automated.
23:23So I'll check-in with you guys when we get those back. So I thought while this is running, and apologies if you hear thunder or fire trucks, I wasn't kidding about the tornado. I thought that I would take a quick second to talk about skills.
23:36Because what happened here is one of them came back with being restricted, just like we saw earlier.
23:43And so what we can do is start to build a bit of a knowledge bank around what prompts get restricted and why and what don't so that it doesn't happen in the future. So what is a skill? A skill is essentially a recipe for an AI agent.
23:55So if someone said, hey, can you make me some chocolate chip pancakes? You would pull up a recipe of chocolate chip pancakes and you would make it. And the next time, you would pull up that same recipe and you would make the pancakes and they'd be the exact same.
24:06But if you didn't have a recipe and you were kind of guessing the measurements and guessing the order and the temperature, your pancakes would come out different every single time. So when we give our agent a skill, it basically means, okay, whenever I want an Instagram ad, you do it exactly like this and now everything feels on brand, everything feels consistent.
24:21So if we bake into a skill, hey, by the way, in the past, you used these five phrases and these five words and it got, you know, basically flagged, so don't ever use those five words or phrases ever again. And the coolest part about skills is every single time that you run your skill, that skill gets better because you can run your skill and you can say, okay.
24:40You just created me these five advertisements with this skill. I don't like x, y, and z, but I love a, b, and c. So update the skill to make sure that next time, you know, it's better.
24:49So while this is actually finishing up and while those generations are happening, let me just show you something. I'm opening up a new chat in this project, and I'm gonna go to our Claude over here.
24:59Let's say that we wanna find one that we really, really like. So, actually, one of my favorite generations so far was this one. What I can do is I can take this prompt and I can copy this prompt.
25:09And essentially, what I'm gonna do is reverse engineer a skill from this prompt. So I'm gonna go back into the Claude.
25:16I'm gonna paste in that prompt, and I'm just gonna start yapping. Hey, Claude. So this prompt that you're looking at right above is my favorite output we've gotten from Higgs Field Marketing Studio.
25:28This was a hyper motion, fast paced kind of like launch video for our product, and I loved it. It had fast cuts.
25:35It had nice zooms. It had nice details. And I want to turn this into a skill that lives locally inside of this project in the dot cloud slash skills so that anytime I ask for a hyper motion style video, you will utilize this and they're always consistent and they always have this style.
25:54So turn this into a skill for me.
25:58And that's basically it. I mean, obviously, the skill's not gonna be perfect on the first shot, but when it comes to actually creating the skill, that's all it takes. And usually the way that I like to build them is I like to play around with a bunch of outputs, you know, maybe generate five different things, and then I pick the one that I like the most or the two or three I like the most and say, how can I reverse engineer a skill from these outputs?
26:18And that's exactly what you just saw me do. Okay. So these should be done now.
26:22I'm gonna open up the sheets and we're gonna take a look Hopefully now, the most important thing I'm really looking for here is that the images or sorry. The picture of our product looks exactly as it should.
26:32So perfect. This one looks exactly like the reference image. This says eighteen months, no sleep, then seven nights of this.
26:38Try it tonight, sixty night money back guarantee. And this looks like an Instagram story style ad. According to the research that it did, this one is kind of what happens and what should be converting.
26:48Let's take a look at this next one here. We have the sleep bottle, which looks exactly as it should, which is perfect. It says melatonin does not equal sleep.
26:55Try the formula. 28,000 parents swear by free shipping. So once again, this looks like it's supposed to be sort of an Instagram story style of advertisement.
27:05Let's take a look at the video we've got here.
27:12Okay. So super short five second video. I don't love this one.
27:16I don't know if it's super engaging, but, hey, this was based on the research that our agent did, and obviously we would test this one. If it's not spending well, we would kill it. This one is more of a square style, so it's looks like it's only one out of five different carousel slides, but why am I exhausted at 2PM even when I sleep?
27:31Eight hours. Advanced sleep formulas, have to see why. Okay.
27:35So that was just really proof of the bottles actually coming through now as they should. This one has a huge stat up top. So we can see that all of these were based on the on screen text, based on the CTA, based on the platform fit and the notes, and all of this other stuff that the agent did research on and helped us ideate on.
27:52Now what I wanna do is let's try to make another one based on this skill. So once this skill has finished up, we will try to invoke it, and we will try to create another hyper motion style video from this skill right here.
28:03Alright. So it created us one called hyper motion video. Now keep in mind, when we installed our Higgs field CLI, we also installed the agent skills.
28:13So even if you don't have this one and you say, hey. Can you create me an image? Can you create me a video?
28:16It should automatically be triggering these other default Higgs field skills automatically. But, anyways, what we're gonna do is we need to open up a new session to try to use this skill.
28:25Okay. Now before I actually use the skill, I wanna check, did it save our reference image or not? So I'm going to go back into this.
28:33I'm gonna click on files, and I wanna see if we have any brand assets. It doesn't look like we do. We have a bunch of data.
28:39Actually, wait. No. We do.
28:40Right here. Data, assets. We have our sleep bottle image.
28:43So you can't see it in this preview, but if I actually open up our folder, which this project lives in, which is right here at Higgs Field Studio, if I go into data and if I go into assets, you can see that it did upload this as a reference image, which is great because what that means is in a new chat, I can tag it.
28:58So it was called sleep bottle reference PNG, so I can use that. And then I can hopefully do a slash command for Higgs field.
29:06Oh, no. It wasn't called Higgs field. It was called hypermotion.
29:09So if I do slash hypermotion okay. I'm not seeing it right now.
29:12Let's try to just invoke it by natural language. So I want you to use the HyperMotion skill and turn this sleep bottle image, which is the reference image I've tagged, and I want you to create a HyperMotion style video in Hicksfield using their marketing studio.
29:28So that's all I'm gonna say for a prompt. I'm not gonna give it anything else. What I'm watching for here is to make sure that it actually invokes the skill because every time Cloud Code invokes a skill, it will tell you that it does.
29:38So hopefully what we see is that it's searching for one and that it actually calls it. If it doesn't, I'm gonna stop this generation. And right now it's running the wrong skill.
29:45It's running Higgs field generate. So this isn't the one we actually wanted to do. We wanted it to run the hyper motion skill.
29:51So what's happening is probably just because we just created that, so it hasn't registered yet. So I'm gonna go ahead and stop this session real quick.
29:58And what we wanna do is check if that skill actually exists. So I'm gonna go to our project. I'm gonna go back to the main section.
30:05The skills will live in a dot Claude folder. So dot Claude, skills, HyperMotion video. The skill does indeed exist, so I'm not exactly sure why it didn't get called.
30:14Let's open it up real quick and take a look at it. This is what a skill file actually looks like. It's just marked down.
30:18It's called HyperMotion video. Generate a HyperMotion style premium product launch video via Higgs Field Marketing Studio, high energy, when to invoke, what to do, what to ask before generating, basically the template.
30:30So here are the hard rules for the skill. So I might just have to, like, close out of the Clot app, open it back up, and then we should be able to see the skill. Alright.
30:37So I close out of the app, and now that I'm back in, I can see the HyperMotion video skill, and I'm just gonna say, this is the skill that I want you to actually use. I also realized that my dictation accidentally corrected HyperMotion to Remotion, so that's why maybe Clog got a little bit confused.
30:53But it's gonna look at the image, it's gonna read the skill, and then we should have a pretty solid output. And because it read the skill, it said, okay. One quick question before I lock in the prompt.
31:03Do you want a model in the ad? Do you want it to be UGC, or do you just want it to be product only? I'm gonna say product only, and hopefully, it should be good to go now.
31:12Alright. So moment of truth. This is done.
31:14I'm gonna go ahead and open it up. Let's see what we got.
31:31Wow. I mean, that's crazy good. I think, you know, obviously, this is where I'm like, okay.
31:37That's pretty bad. But all of this stuff, I mean, like, it followed the skill and we'd be able to say, hey.
31:43I like this. I don't like this, and you make it a little bit better. But I think that this is really good.
31:48I mean, I really liked the feel. It looks very real. And, yeah, I am pretty happy with this output.
31:54Now it is a little bit unfortunate because in the reference image, all of the words do come through perfect. So it's not an issue with the reference image.
32:02It's an issue with the model that generated this video. Sometimes models are going to mess up text a little bit, especially when you're doing image to video rather than image to image.
32:13Now one thing to keep in mind is this is the . Every day, every month, they're going to get better and better. So right now is the worst it'll ever be.
32:22But I do wonder if there's some stuff that you could do in that video and in the prompting of the video to actually make it a little bit more accurate. And I think if it were me and I had that output and I just, you know, was getting that sort of quality but the words weren't coming through right, then then I would probably just say, you know what?
32:37That's fine. For our videos that are this high quality, we're just going to use a different label cover, which is just the logo and the name rather than like all the little metadata that we might not need right now. So anyways, there there are some ways to work around.
32:49There's some ways to have a better positive attitude about it rather than just hating on it and calling everything AI slob. And because this was using marketing studio video, it said that it didn't actually use specifically one model.
33:00It kind of used a mix because it used, you know, marketing studio. So you don't choose it directly unless you would have ran it with some sort of flag. So anyways, this is kind of the main workflow.
33:10Right? We're using it to ideate. We're filling in a sheet.
33:12We're marking off statuses. But how do we actually start to actually automate this? Well, think about this.
33:18In Claude, we're able to set up these things called routines. And what routines are is it basically just, on a set cadence, injects a prompt into Claude code. So what we just did here, this could be a routine.
33:29All we have to do is prompt it to do so. So in a routine, what we could do is we could build a new one that basically says, okay. Every Sunday, I want you to look at this Google Sheet, and I want you to also pull in data from, you know, Instagram or wherever we're posting this stuff, and I want you to analyze what's working, what's not, and then I want you to ideate and I want you to add on top of the sheet 50 new generations.
33:50So 50 new every single week. And then what we do is that let's say that that's a Sunday night. We could maybe have a Monday morning one, that's a new routine.
33:59Okay. Monday morning, you're gonna go to the sheet and you're gonna pick 30 videos with a blank status, and that's how we ensure that we're not duplicating efforts. So let's say it picks, you know, these 30.
34:08It's not 30, but let's say it picks all of these because there's no actual status. It would grab all 30 of these. It would create the prompts for all of them.
34:17It would generate all of them, and then we'd wake up on Monday morning with all of these completed with URLs and with job IDs. And then we could scale it up. Maybe we're doing planning every Sunday and Thursday, and then maybe we're doing generating every Monday and Friday, and maybe we're scaling up from 50 to a 100 or 200.
34:33And that's how you sort of push the system to the point where it's scaling way faster than you could as a human or even maybe multiple humans. Then if you wanna take it one step further, you could actually connect this pipeline once you actually get it in a place where you trust the outputs to something like Potato or even maybe plug it into the meta ads manager.
34:49And you're starting to schedule and post these things automatically too because you've built up a batch of skills that you actually trust enough to just let them run autonomously. So that's kind of the idea. And everyone's gonna get in here and everyone's gonna do it a little bit differently, but it's really, really simple to set up your routines right in here or to come into here and say, hey.
35:05I want you to set up a routine for me, 8AM, Monday. Do this and this Because it's all very possible with natural language. So if you guys wanna dive a little bit deeper into routines, I'll tag a full video right here where I dive into how you set them up and some of the little gotchas that are in there.
35:19But that is gonna do it for this video. So if you guys enjoyed, please give it a like. It helps me out a ton.
35:22And as always, I appreciate you guys making it to the end of the video, and I will see you in the next one. Thanks, ever
The Hook
The bait, then the rug-pull.
Nate Herk opens with the promise stated as a fait accompli: Higgsfield has every state-of-the-art image and video model, Claude knows how to talk to it, and together they ideate and generate 'a 100 times faster than the average human could.' Then he immediately cuts to a wall of finished ads — headphone halos, sleep-supplement bottles, hyper-motion launch videos — generated 'in literally five minutes with one prompt.' Promise + proof, in the first thirty seconds.
Named ideas worth stealing.
10:45concept
MCP vs CLI for Agents
MCP exposes every tool by default — token-heavy. The CLI is the lean alternative: same capabilities, faster, cheaper, better for agents that loop. Default to CLI when wiring tools into Claude Code; reserve MCP for ad-hoc claude.ai exploration.
Steal forany agent project where token cost compounds across iterations — Joe's Trigger.dev tasks, Paperclip orchestration
23:50concept
Skill = Recipe (pancake analogy)
A skill is to an AI agent what a recipe is to a cook: lock the inputs, ordering, and constraints so output stays consistent run-over-run. Bake in negative constraints too (words that triggered moderation flags, brand assets that must lock).
Steal forJoe's .claude/skills layer in MCN — codify Killing Excuses shot recipes, Mod Producer runsheet templates, ad creative voice rules
25:05model
Reverse-engineer a skill from a winning prompt
- Generate 5+ variants of a creative
- Pick the 1–3 outputs you actually love
- Copy the exact prompt that produced them
- Paste back into a fresh chat and say: 'turn this into a skill in .claude/skills so anytime I ask for X, use this'
- Iterate the skill every run — tell it what you liked/didn't and have it self-update
Don’t write skills from scratch — mine them from outputs that already worked. Outputs first, recipes second.
Steal forevery winning ad / video / email in MCN should be reverse-engineered into a skill the next agent run can replay
12:18concept
Bring outside expertise into the project as Markdown
Don’t expect the base model to be a master copywriter. Run a deep-research prompt against Twitter threads / YouTube / Perplexity / books, save the result as `advertising_masterclass.md` in the project, then @-tag it whenever you ideate. The model now has subject-matter expertise on tap.
Steal forJoe's Maria Wendt 255-email swipe file, HTSS notes, Hoffmann's Reward Funnel — land them as project-local .md files for every MCN content agent
15:20model
Master Sheet + status column = creative ops database
- Tabs: generations, by-product, by-style, planning, creative slate
- Columns: product, style, image/video, model, prompt, status, result URL, job ID
- Status drives which rows the next routine picks up
- Sheet is queryable by the agent for analysis later
One source of truth makes the rest of the automation possible. The status column is the glue — it’s what stops agents from duplicating work across routine runs.
Steal forJoe's Content Forge meme/quote pipeline, Mod Producer runsheets — a Sheet with a status column beats any custom DB for v1
33:45model
Routines: ideate Sunday → generate Monday
- Sunday 9pm — 'analyze sheet + platform data, add 50 new generation ideas with empty status'
- Monday 8am — 'pick 30 blank-status rows, generate prompts + assets, mark complete'
- Add Thursday/Friday as a second cycle to double throughput
- Optional: pipe completed assets to Potato or Meta Ads Manager for posting
Two cron-style prompts and you have an autonomous ad-generation loop. The cadence (Sun plan / Mon execute) prevents the agent from generating against stale ideas.
Steal forJACE / REESE routines on the MCN side — weekly content planning + execution cycles for Killing Excuses and Notes to Myself
How they asked for the click.
VERBAL ASK
34:59next-video
“If you guys wanna dive a little bit deeper into routines, I'll tag a full video right here where I dive into how you set them up… if you guys enjoyed, please give it a like.”
Soft CTA — anchors on the next-video tag for routines and a like ask. No newsletter push, no product (other than the embedded Glido sponsor at 20:33). Trade-off: zero conversion pressure, but also no lead capture from a 35-min watch.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
AFFILIATECommission earned if you click.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
