Baue & automatisiere dein Marketing mit Claude Code — Transcript
Full transcript
- 0:00This here is my marketing hub, and right
- 0:02now three agents are working
- 0:03simultaneously. One is building a page,
- 0:06one is checking it against 53 rules, and
- 0:08one is currently designing new ads.
- 0:10Every morning at 9:00 AM, I get a report
- 0:12on every channel.
- 0:14It shows how many clicks and leads there
- 0:16were yesterday and which search terms
- 0:18burned money, just to continuously
- 0:19optimize the ads and see what is working
- 0:21well and what needs changing.
- 0:24Every lead lands here with the channel,
- 0:26the campaign, and the search term it
- 0:27ultimately came from.
- 0:30And if I reject any of this work, the
- 0:32agent team turns it into a fixed rule so
- 0:33that this mistake doesn't happen again.
- 0:36Behind all of this are 1,404 sessions
- 0:38with Claude and Codex, and in this video
- 0:40, I'll first show you what this hub
- 0:42actually is and what it can do.
- 0:44After that, we'll build it together,
- 0:46prompt by prompt, so that in the end,
- 0:48you have your own marketing hub tailored
- 0:50to your company.
- 0:51Let's start with the overview, because
- 0:53here I see at a glance every morning
- 0:55where my marketing stands. As an example
- 0:57, the company Blauwerk Pools is running
- 0:59here, a pool builder from Hanover.
- 1:01I completely made this company up for
- 1:03this video, but I have applied the
- 1:04principles and everything you see here
- 1:06multiple times exactly like this and
- 1:07used it for myself as well.
- 1:09And the numbers here are sample figures
- 1:11, but they aren't completely pulled out
- 1:12of thin air either. At the very top are
- 1:14the four figures that really count for a
- 1:16business: visitors, leads, the
- 1:17conversion rate between them, and the
- 1:19money currently outstanding in quotes.
- 1:21And below that, you see each channel
- 1:23individually: Google Ads, Meta, and
- 1:25organic search, each with visitors,
- 1:27leads, and the respective conversion
- 1:29rate. This way, I see immediately which
- 1:31channel brings leads and which is
- 1:32currently only bringing clicks. On the
- 1:34right is everything that is waiting for
- 1:35my decision. Right now, that’s a new
- 1:37landing page. The check has already
- 1:39verified it against three checklists,
- 1:41and I can now accept it, reject it, or
- 1:43ask questions about what exactly it is
- 1:45or whatever else is still on my mind
- 1:48that isn't clearly defined. When I click
- 1:50on "ask," Claude opens with everything
- 1:52belonging to this page, and I can ask
- 1:54directly why something is built the way
- 1:56it is. For example, I ask why the form
- 1:58is on the right and not at the bottom.
- 1:59Claude provides the rule behind it and
- 2:02the test that proves it. At the bottom,
- 2:04you can see which agents are currently
- 2:05working and on what. Next to it, the
- 2:07notifications come in, like the daily
- 2:09Google Ads report or a new inquiry that
- 2:11just arrived. This way, you always have
- 2:13a live overview, and on the left here is
- 2:15my actual activity. Each box is a day,
- 2:17and the brighter it is, the more
- 2:19sessions I had that day with Claude Code
- 2:21or Codex. The most important area for
- 2:23the business is the inquiries, because
- 2:24that is what marketing is ultimately
- 2:26measured by. Every inquiry here arrives
- 2:28as a card. It states what the customer
- 2:30wants, e.g., an 8 x 4 m concrete pool,
- 2:33where they live, and roughly how big the
- 2:35project is. Below that, it says where
- 2:36the inquiry came from. This is Google
- 2:38Ads, and the card even shows the search
- 2:40term the customer used to find us. I’m
- 2:42now sending a test inquiry via the form
- 2:44on the landing page. A few seconds later
- 2:46, it’s there. The office has already
- 2:48sent the customer a confirmation, and
- 2:50management has scheduled a callback.
- 2:52Once I’ve spoken with the customer, I
- 2:54simply drag the card to the next column
- 2:55. From the site visit, it moves to the
- 2:57quote and then to the order.
- 3:00With every won order, the hub reports
- 3:02back to Google Ads which search term
- 3:04ultimately brought in revenue, allowing
- 3:05Google's system and its AI to further
- 3:07optimize.
- 3:09Now for the reports, because this is
- 3:11where the hub works without me having to
- 3:13do anything. Every morning at 9:00 AM, a
- 3:15daily check runs for every channel. For
- 3:17Google Ads, it reads all of yesterday’s
- 3:19metrics—impressions, clicks, costs, and
- 3:21inquiries. Then it goes through every
- 3:23search term that someone clicked on.
- 3:25Terms that are clearly unrelated to the
- 3:27offer are immediately excluded. This
- 3:29provides further optimization, such as
- 3:31job searches, DIY instructions, or the
- 3:33names of other companies—competitors—if
- 3:35you don't want them. Anything that isn't
- 3:38clear is not blocked. That ends up here
- 3:40under review, where I decide myself, and
- 3:41at the end, there is a recommendation
- 3:43with the relevant figures.
- 3:45Here, it suggests directing the search
- 3:46term "concrete pool price" to its own
- 3:48page because price-seekers aren't
- 3:49inquiring on the pool construction page
- 3:51.
- 3:51This control is not just a gimmick. A
- 3:53daily check like this has been running
- 3:55for the Google Ads account every morning
- 3:57at 9:00 AM since October. Budgets, bids
- 3:59, and ads are never changed; that stays
- 4:01the same, and on Mondays, the weekly
- 4:03analysis follows. An agent compares all
- 4:05channels with one another and writes up
- 4:06recommendations for me based on that.
- 4:08For instance, it found that the hero
- 4:10section with the form on the right
- 4:11brings in significantly more inquiries
- 4:13than the old version. Additionally, it
- 4:15reports that a creative on Meta is
- 4:17fatiguing because the target audience
- 4:18has already seen it over four times. I
- 4:20can assign any recommendation to the
- 4:22appropriate agent with a single click.
- 4:24Here, the task goes to the ads team, and
- 4:26they immediately start on three new
- 4:27motifs.
- 4:28To ensure the agents don't just produce
- 4:30anything, the headquarters knows the
- 4:32entire funnel.
- 4:33Here you can see every stage from first
- 4:35contact to the order. Between the stages
- 4:37, it shows how many people move forward
- 4:39, i.e., the percentage conversion rate.
- 4:41And the headquarters directly marks the
- 4:43bottleneck, showing exactly where the
- 4:45funnel is stalling. Here it is the step
- 4:47from visitor to inquiry, which turns
- 4:48directly into a task for the website.
- 4:50Below that is the strategy for each
- 4:52stage. Those who haven't even thought
- 4:54about a pool yet get a before-and-after
- 4:55Reel. Those who are already comparing
- 4:57prices get a guide on the costs. And
- 4:59those who want to make an inquiry now
- 5:01land on the landing page with the form.
- 5:03To do this, the headquarters knows the
- 5:04target groups. For each group, it lists
- 5:06what they desire, what they are looking
- 5:08for, and what makes them hesitate. For
- 5:11every objection, there is practically a
- 5:12matching response. And at the bottom,
- 5:14you can see immediately where content is
- 5:15still missing. For couples over 50, for
- 5:17example, there is nothing yet for the
- 5:18initial contact. This also goes out as a
- 5:20task with one click. Everything the
- 5:22agents need to know about the company is
- 5:23in the company profile. This lists the
- 5:25services, the region, and the goal.
- 5:27Below that, it states for each channel
- 5:28who is allowed to do what. On the
- 5:29website, the team may create drafts but
- 5:31cannot publish anything. For Google Ads
- 5:33and Meta, every change ultimately needs
- 5:36my approval, as does every budget. And
- 5:38here it shows which claims are
- 5:39substantiated and which are still
- 5:40missing. Blauwerk Pools does not have
- 5:42reviews, reference projects, or
- 5:44certificates yet. That is why no agent
- 5:46writes such things on a page. It places
- 5:48a visible placeholder there until real
- 5:50data is available. At the very bottom,
- 5:52you see the integrations, meaning where
- 5:53the numbers come from: Google Ads,
- 5:54Search Console, Meta, the form, and
- 5:56keyword data. Now to an area that saves
- 5:58me a lot of work: the Reel Lab. For
- 6:00example, I simply enter an Instagram
- 6:02profile here, such as a strong profile
- 6:04from the industry, and click analyze.
- 6:06The headquarters pulls the last 30 Reels
- 6:08, analyzes every image and sound, and
- 6:10rates the profile. Here you can see the
- 6:12strongest Reels, and for each one, it
- 6:13shows how far it exceeds the average of
- 6:15this profile. On the right, it shows
- 6:17what works: before and after in the
- 6:18first few seconds, a length of 9 to 14
- 6:20seconds, and the price in the title.
- 6:22From this, the hub creates a video idea
- 6:24for your company, already broken down
- 6:26into scenes with seconds. With one click
- 6:28, it goes to the video agent. In the
- 6:29YouTube section, things continue in the
- 6:31same way. Each video idea moves here
- 6:33from the concept through the script and
- 6:35filming to the finished video. For every
- 6:37video, it shows how often it was viewed
- 6:38, how many people clicked on the landing
- 6:40page afterward, and how many inquiries
- 6:42resulted from it. The ideas come from
- 6:44three sources: from search terms, from
- 6:46the Real Lab, and from questions
- 6:47customers ask in their inquiries. In the
- 6:49SEO section, I can see how the company
- 6:51ranks on Google and in AI search. Here
- 6:53are the search terms with search volume
- 6:54and position, and next to them, the
- 6:56trend over the last few weeks. Below
- 6:57that, the hub checks whether the company
- 6:59is mentioned in AI answers. To do this,
- 7:01it asks ten typical questions in ChatGPT
- 7:03, Gemini, and Google's AI overview, and
- 7:05the SEO reports are kept here. Each
- 7:07report has action items, and each item
- 7:09goes directly to an agent. This is the
- 7:10actual audit report for the landing page
- 7:12. 34 out of 52 rules are met, and the
- 7:14open points can only be completed at
- 7:16launch. For example, HTTPS and the
- 7:19sitemap. Lighthouse shows 98 on mobile,
- 7:21so that's the performance. Now I'll show
- 7:23you how the team processes a task from
- 7:25start to finish. This is a real order
- 7:28that I just placed in the hub. Under
- 7:30websites, you'll find the landing page
- 7:31for Blauwerk Pulse. Here you can see all
- 7:33three versions side-by-side, and here is
- 7:35the entire process. Management accepts
- 7:37the order and plans the steps. The
- 7:39research team estimates the costs for
- 7:41the keyword query and pauses until I
- 7:43approve it. After that, we start with
- 7:44SEO according to our specifications. The
- 7:46copy for the page is written, and the
- 7:48website is built. The audit finds nine
- 7:50errors in the first round and sends the
- 7:51page back. I then rejected this version
- 7:53. My reason was a white veil over the
- 7:55photo, which made the whole page look
- 7:57pale. The team immediately turned that
- 7:59into a rule. The second version was
- 8:00better, but it still felt like an
- 8:02AI-generated page to me, with too much
- 8:04text and the form only at the very
- 8:05bottom. So I rejected that one as well.
- 8:07Three new rules emerged from this, and I
- 8:09just accepted the third version. A short
- 8:12hero, a large photo, and the form right
- 8:14on the right in the first frame. These
- 8:16rules end up here under "learned," and
- 8:18that is the part that makes the hub
- 8:19better every week. A new rule does not
- 8:21apply immediately. It waits for my
- 8:23go-ahead first. After that, it is on
- 8:24probation and must prove itself through
- 8:26three tasks. Only when the same mistake
- 8:27hasn't occurred three times in a row
- 8:29does the rule become fixed for every new
- 8:31task. And here is the team that does all
- 8:33the work. Ten agents, each with a clear
- 8:35task. For each one, you can see what
- 8:37it's currently working on, how often it
- 8:39has run, how many of its results I've
- 8:40accepted, and what rules it has learned
- 8:41. And every agent has a limit it never
- 8:44crosses. The ads never change a budget,
- 8:46the websites never go live without my
- 8:48go-ahead, and the research tools don't
- 8:50pull anything that costs money before
- 8:52I've at least approved it. Under
- 8:54workflows are the fixed paths through
- 8:56the team. Every workflow ends with me,
- 8:58and next to it are the templates,
- 8:59meaning how a finished result must look
- 9:02. Next to that are the skills, the
- 9:03knowledge base, and the checklists. The
- 9:05skills are the agents 'tools, for
- 9:06example, the SEO audit or the skill for
- 9:08sales-optimized websites. The knowledge
- 9:10base contains every rule with its source
- 9:12, for example, directly from Google's
- 9:13documentation. And from this knowledge,
- 9:15the checklists are built, against which
- 9:17the audit checks every page. For the
- 9:19landing page, you can see here for every
- 9:20point whether it is met, only partially
- 9:22met, or still completely missing. So,
- 9:24now let's build this headquarters
- 9:26together. I will now show you the prompt
- 9:28I enter into Claude Code and the result
- 9:29it produces. For this, you only need
- 9:31Claude Code and an empty folder on your
- 9:33computer. In the first step, we set up
- 9:34the basic structure. I open Claude Code
- 9:36in an empty folder and enter this prompt
- 9:38. Create a marketing headquarters for my
- 9:40company in this folder.
- 9:42Every role, every piece of knowledge,
- 9:43every checklist, and every result is its
- 9:45own text file.
- 9:46Write a Claude.md that explains how the
- 9:48team works. Nothing is published and no
- 9:50money is spent before I have approved it
- 9:52.
- 9:53Claude is now creating the folders for
- 9:55approvals, requests, agents, skills,
- 9:57knowledge, checklists, reports, and also
- 9:59lessons learned. And it writes the
- 10:01Claude.md.
- 10:02Claude reads this file first at every
- 10:04session, and it contains how the team
- 10:05works and which rules always apply.
- 10:07Let's take a quick look inside. It
- 10:08states that every assignment first goes
- 10:10to management. Management plans the
- 10:11steps and distributes them to the agents
- 10:13. Below that, it states that every
- 10:15result is checked before it is submitted
- 10:17for approval and that every step ends up
- 10:19in the log with a timestamp, agent, and
- 10:21file. That is basically the operating
- 10:22manual for your entire team, and you can
- 10:24change it yourself at any time. In the
- 10:26second step, the headquarters gets
- 10:27everything about the company. For this,
- 10:28I simply enter the company data directly
- 10:30into Claude. Here is an example of my
- 10:32company data: Blau Werk Pools, pool
- 10:33construction in Hanover and the
- 10:34surrounding area. We build concrete
- 10:36pools and prefabricated basins, renovate
- 10:38old pools, and handle roofing and
- 10:40equipment. The goal is more inquiries
- 10:42for new pools; create the company
- 10:44profile and target groups from this and
- 10:46ask me what is still missing. Claude
- 10:48creates the company profile and derives
- 10:49three target groups from it, each with
- 10:51wishes, search terms, and objections.
- 10:53And then it asks about what it cannot
- 10:55know. For example, whether there are
- 10:56genuine reviews, what reference projects
- 10:58exist, and how long the company has been
- 11:00in business. I answer what I know
- 11:01directly in the chat. For example, that
- 11:03most customers are families with gardens
- 11:05and that many ask about the costs first
- 11:06. Claude immediately enters this into
- 11:08the company profile and the appropriate
- 11:09target group. Everything I have not yet
- 11:11answered remains empty. No agent will
- 11:14later write a review or a figure on a
- 11:16page that is not verified. In the third
- 11:18step, we set up the team. Create ten
- 11:20agents: Management, Research, SEO,
- 11:22Websites, Copywriting, Ads, Video,
- 11:23Quality Assurance, Secretariat, and
- 11:25Learning.
- 11:27Everyone gets a task, their skills,
- 11:28their knowledge, who they hand over to,
- 11:30and a clear boundary they never cross.
- 11:33A separate file is created for each
- 11:35agent. Here you can see the SEO agent's
- 11:37file. At the top it says what they do,
- 11:38below that which skills they use, which
- 11:40knowledge they read, who they hand over
- 11:42to, and what they never do. Now the
- 11:43agents get their tools, the skills. A
- 11:45skill is a finished set of instructions
- 11:47according to which Claude always
- 11:49completes a specific job in the same way
- 11:50. Connect the agents with my skills. The
- 11:53SEO agent uses the skill for SEO audits
- 11:55and keyword research. The website agent
- 11:57uses the skill for sales-optimized pages
- 11:59, and the ad agent uses the skill for ad
- 12:01creatives.
- 12:02Claude notes for each agent which skills
- 12:04they use and creates a short card for
- 12:05each skill.
- 12:07On it is the command, what the result
- 12:08should be, and whether the skill costs
- 12:10anything. If you don't have your own
- 12:12skills yet, that's no problem. You can
- 12:13simply ask Claude to build a skill from
- 12:16a task you already do often anyway. So
- 12:18Claude can analyze and build that for
- 12:19you. The next time, it will then run
- 12:21according to the same instructions. In
- 12:22the fourth step, the team gets its
- 12:24knowledge; build the knowledge base for
- 12:25SEO, AI search, local search, landing
- 12:27pages, and sales psychology. Every rule
- 12:29needs a source, preferably directly from
- 12:31Google. Create checklists from this,
- 12:32against which the Quality Assurance
- 12:33agent checks every result. Claude now
- 12:35researches official sources and writes
- 12:37down every rule along with its source.
- 12:39Behind it is the link to the source at
- 12:41Google. You can verify every rule
- 12:42yourself and the team doesn't invent
- 12:44their own SEO tips. This is where the
- 12:46checklists come from. If the audit
- 12:47rejects a page later, it can always say
- 12:50exactly which point wasn't met and where
- 12:52the rule comes from. In the fifth step,
- 12:54we build in the two rules that make the
- 12:55whole thing secure and adaptive. Every
- 12:57result first lands under approvals and
- 13:00waits for my "yes." Every paid query
- 13:02shows the estimated costs beforehand.
- 13:04When I reject something, I write a
- 13:06learning note and turn it into a rule
- 13:08that only becomes permanent after my
- 13:09approval and three tasks without relapse
- 13:11. Claude writes these rules into the
- 13:13cloud and creates the "learned" folder.
- 13:15From now on, every rejection
- 13:16automatically becomes a rule. This is
- 13:18how it looked for the landing page. I
- 13:19wrote that the page feels like an AI
- 13:21page. The team turned this into a
- 13:23learning note with my reason, the error
- 13:25, and the new rule. The rule is now:
- 13:27short hero section, large photo, and the
- 13:29form on the right. And because it is
- 13:31registered under websites, the agency
- 13:32reads it before every new page. In the
- 13:34sixth step, we connect the headquarters
- 13:36with the channels. Connect my inquiry
- 13:38form, Google Ads, and the Search Console
- 13:40. Every new inquiry arrives with channel
- 13:42, campaign, and search term in the
- 13:43inquiries folder. Only read Google Ads
- 13:45and Search Console and never change
- 13:47anything there without my approval.
- 13:49Claude writes small scripts for this.
- 13:50The form now sends every inquiry as a
- 13:52file to the folder and the ad data is
- 13:54only read daily. This is what an inquiry
- 13:56file looks like. At the top are name,
- 13:58location, and project, below that the
- 14:00channel, the campaign, and the search
- 14:02term. The board turns this into the card
- 14:04you saw earlier in the Kanban. For the
- 14:06connection to Google, you need your
- 14:07access credentials once. Claude explains
- 14:09every step to you, and the keys are then
- 14:10stored securely in your keychain and not
- 14:12in the folder. In the seventh step, we
- 14:14let the headquarters work by itself. Set
- 14:16up a daily report for every channel at
- 14:189:00 AM each morning. You may exclude
- 14:20clearly incorrect search terms.
- 14:22Everything else you present to me for
- 14:24review, and every Monday, a weekly
- 14:25analysis should compare all channels and
- 14:27give its recommendation as an order to
- 14:29the agents. Claude creates the orders
- 14:31for you and schedules them. Each order
- 14:33states exactly what the controller is
- 14:34allowed to do and what not. It is only
- 14:36allowed to exclude terms that clearly do
- 14:38not fit, such as job hunting, DIY guides
- 14:40, material purchases without tradesmen,
- 14:42or third-party company names. General
- 14:44terms like pool construction, car
- 14:45dealership in the catchment area, and
- 14:47anything that has generated an inquiry
- 14:48before, it never blocks. These
- 14:50boundaries are important, otherwise an
- 14:52agent might eventually block a term that
- 14:53actually brings in orders. Let's build
- 14:55the board, meaning the interface you've
- 14:57been seeing the whole time. Let's build
- 14:59a board that reads this folder: the
- 15:00areas on the left, key figures at the
- 15:02top, approvals with accept, reject,
- 15:04query options, and inquiries as a Kanban
- 15:06board, in black, silver, and metallic.
- 15:08Claude is now building a small website
- 15:10that runs directly on your computer. It
- 15:12reads the folder, and you can
- 15:13immediately see every file that an agent
- 15:14creates or modifies here. Of course,
- 15:16your design won't look exactly the same
- 15:18right away; that's a matter of taste,
- 15:20naturally. You can design the whole
- 15:21thing however you want it to be. This is
- 15:23my design for now. When you click on "
- 15:25Accept" here, the board writes the
- 15:26decision back to the folder and the team
- 15:28continues working. The board wasn't
- 15:30finished like this on my first try
- 15:32either. I told Claude what I didn't like
- 15:34, for example, that the numbers were too
- 15:35small or that I wanted the channels as a
- 15:37table. Claude then changes it directly,
- 15:39and after a few hours, it looks like it
- 15:41does here for me, tailored to your taste
- 15:43and your design principles, of course.
- 15:45And now comes the best part about the
- 15:46entire hub. Behind everything you've
- 15:48seen in this video is just a folder with
- 15:50text files.
- 15:51The agents, the knowledge, the
- 15:53checklists, the inquiries, and the rules
- 15:54are all normal files that you can read
- 15:56and change.
- 15:57So, you don't need to install anything
- 15:59or buy expensive software.
- 16:01And because all the knowledge is in the
- 16:03folder and not in the model, you can
- 16:04also give the same folder to Codex later
- 16:06, for example, or any other arbitrary
- 16:08LLM.
- 16:08It reads the same files and works
- 16:10according to the same rules. The longer
- 16:12you work with it, the better the hub
- 16:13becomes. Every rejection turns into a
- 16:15rule, and every week new figures and new
- 16:17recommendations are added.
- 16:19If you want to build a hub like this, or
- 16:21simply download it, and want to learn
- 16:22more about AI marketing, feel free to
- 16:24join our AI Marketer community.
- 16:26You can find all further details there.
- 16:28The link is in the video description,
- 16:30and you can also simply scan the QR code
- 16:32in the top left or right.
About this transcript
This page contains the full transcript of Baue & automatisiere dein Marketing mit Claude Code by Ismajl Pukaj | KI Agentur, generated from the public captions YouTube serves with the video. The transcript has 3,686 words across 550 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.