6:02:49Nick Saraev · Tutorial
Claude Code for Marketing: The Full Course
Six hours of screen-share in which one operator rebuilds an entire marketing department out of prompts, skills, loops and scheduled cloud routines.
August 8thThe argument in one line.
A five-step audit, plan, eval, skill, and schedule process, run inside Claude Code for one week, can automate the large majority of a knowledge-work business's non-client-facing tasks, leaving only a handful of human-judgment tasks behind.
Read if. Skip if.
READ IF YOU ARE…
- You run a service business, agency, or any knowledge-work operation and want a concrete method for finding what to automate first.
- You already use Claude Code (or a similar AI agent) for work tasks and want a repeatable way to turn one-off prompts into dependable, shareable systems.
- You manage a team and want non-technical people to benefit from AI automation without learning to code.
- You're deciding whether an AI agent can really run unattended parts of your business, not just answer questions.
SKIP IF…
- You want a tool-by-tool tutorial. This video describes a five-step process, not specific prompts or screen-by-screen walkthroughs (those are in the linked free prompt pack).
- You're looking for automation inside a technical engineering or DevOps workflow. The examples here are agency operations like lead gen, sales, and client management.
The full version, fast.
The process runs in five steps. First, audit a full week of work: Claude screenshots your screen every few seconds and periodically asks what you're doing and why, producing a complete list of every task in the business. Second, feed that list to a new Claude session and describe each task by voice so it can design an automated system for each one. Third, build evals: run each system ten times, score the outputs out of ten, and loop with Claude until most systems score 7 or 8 or higher. Fourth, turn every system that passes into a reusable skill non-technical teammates can run too. Finally, decide which skills can run on a schedule with no human involved and which still need a person reviewing the output first.
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 – 00:30
01 · I Let Claude Run My Agency
Cold open: Claude Code automated a real million-dollar-a-year agency. Framed as not clickbait, four years of prior disappointment with earlier models.
00:30 – 01:29
02 · Audit Your Work
A week-long audit: Claude screenshots the screen roughly every five seconds and periodically asks what you're working on and why, building a database of the real workday.
01:29 – 02:02
03 · List Every Task
The audit produces a full written list of every deliverable and process touched in the agency, grouped by lead gen, sales, client management, onboarding, and maintenance.
02:02 – 02:33
04 · Have Claude Plan the Systems
Feed the task list to a fresh Claude session, ask it to design automation for each task, and fill context gaps by voice-answering its follow-up questions.
02:33 – 03:58
05 · Build Evals (Aim for 8/10)
Define evals: run each proposed system ten times, score every output, and loop with Claude on specific feedback until most systems land at 7 or 8 out of 10.
03:58 – 04:20
06 · Turn Winners Into Skills
Systems that pass their evals get packaged as reusable skills, which non-technical teammates on a shared plan can run without touching the underlying prompt.
04:20 – 04:41
07 · Batch Your Tests
Run three or four eval-improvement tests in parallel instead of one change at a time, cutting a week of iteration down to roughly a day.
04:41 – 05:34
08 · Schedule What Runs Without You
For each finished skill, decide whether it can run fully on a schedule (a daily client status draft) or needs a human glance first (a full cold email campaign).
05:34 – 06:35
09 · What I Still Do Myself
Roughly 80-90 of 90 daily tasks end up automated. What's left: showing up to calls, diagnosing client problems, and proposing solutions from experience.
06:35 – 07:30
10 · Get the Free Prompts
Advice to use the most frontier model on max thinking for this planning stage, then two CTAs: a free prompt pack (Maker Zero) and a paid 90-day accountability community (Maker School).
Lines worth screenshotting.
- A one-week screenshot audit, captured every few seconds and paired with periodic AI-asked questions, surfaces a business's full task list more accurately than trying to write it from memory.
- Feeding a complete task list into a fresh AI session, then answering its follow-up questions by voice, turns about an hour of talking into enough context to design automated systems for dozens of tasks.
- An eval is a test run ten times on the same task so output quality gets scored and averaged instead of judged from one lucky result.
- Scoring ten outputs per task and only accepting systems that land at 7 or 8 out of 10 turns 'the AI did it once' into something closer to dependable.
- Building and scoring evals, not the monitoring itself, is the most time-intensive part of automating a business this way.
- Turning a passing system into a shareable skill is what lets non-technical teammates get the automation's value without ever touching the underlying AI tool directly.
- Batching three or four eval-improvement tests at once, instead of testing changes one at a time, can cut a week of iteration down to about a day.
- Not every working system should run unattended. Tasks with real stakes, like a full cold email campaign, still get scheduled as human-in-the-loop rather than fully autonomous.
- After a full audit-to-skill cycle, roughly 80 to 90 of around 90 daily tasks ended up fully automated, leaving only five or six for a human.
- Removing nearly all repeatable deliverable work from a job can leave only the parts that were never really about output: sitting with a client to figure out what's actually wrong.
Automate Almost Everything, Then Keep The Rest
AUTOMATION PLAYBOOK
A disciplined audit, eval, and skill loop can turn dozens of daily agency tasks into a handful of human-only ones, as long as every automated system earns its place with a real score, not a demo.
02Audit Your Work
- Before automating anything, spend about a week capturing what you actually do: a screenshot every few seconds plus periodic questions about what you're working on and why.
- Telling an AI what you do from memory misses detail. Showing it a week of real screenshots and in-the-moment answers surfaces the full picture instead.
03List Every Task
- Turn the audit into one explicit list of every deliverable and process you touch, even small recurring ones like checking bookings or sending a welcome email.
- A real knowledge-work task list runs long: in this case it covered lead gen, sales, onboarding, client management, and maintenance, dozens of tasks across five categories.
04Have Claude Plan the Systems
- Hand the full task list to a fresh AI session and ask it to design an automated system for each one, filling context gaps by voice-answering its follow-up questions.
- About an hour of spoken answers was enough context for the system to start designing automation for dozens of separate tasks.
05Build Evals (Aim for 8/10)
- Run each proposed system ten times and score every output yourself before trusting it, since the same prompt can produce a great result once and a poor one the next.
- Treat 7 or 8 out of 10 as the bar for good enough to automate. Below that, keep telling the AI specifically what's wrong and what you want instead until the score climbs.
- Scoring and iterating on evals, not the monitoring step, is the most time-intensive part of the whole process, so budget real hours for it.
06Turn Winners Into Skills
- Once a system consistently scores well, package it as a reusable skill instead of leaving it as a one-off prompt, so it survives past the session that built it.
- Packaging automation as a skill is what lets non-technical teammates use the result without ever having to write or understand the underlying prompt.
07Batch Your Tests
- Test three or four skill improvements at once instead of one change at a time. Running changes in parallel can turn a week of iteration into about a day.
08Schedule What Runs Without You
- Not every working skill should run unattended. Ask, for each one, whether the output needs a human glance before it goes out, and schedule only the ones that don't.
- Low-stakes, repetitive outputs like a daily status draft are good candidates for full automation. Higher-stakes one-shot work like a full campaign is better kept human-in-the-loop.
09What I Still Do Myself
- After automating the deliverable work, what's left is usually the part that was never really about output: talking to people, diagnosing their actual problem, and proposing a solution from experience.
- Going from dozens of daily tasks down to five or six is possible once the systems behind them are proven reliable, not just plausible-sounding in a demo.
10Get the Free Prompts
- Use the most capable model you can, on maximum thinking effort, for the planning stage specifically. It's upstream of everything else, so a weaker plan there weakens every system built from it.
Terms worth knowing.
- Eval
- A test that runs an AI system on the same task multiple times (commonly ten) so its output quality can be scored and averaged, rather than judged from a single result.
- Skill
- A packaged, repeatable AI workflow built once and then reused, and shared with teammates, without rebuilding the prompt or process from scratch each time.
- Max Thinking
- A higher-effort reasoning setting for an AI model that trades more computation time for a more thorough, deliberate answer, useful for complex planning tasks.
- Human in the loop
- A workflow where an AI produces a draft or decision but a person still reviews or approves it before it takes effect, as opposed to running fully unattended.
Things they pointed at.
00:00toolClaude Opus (Max Thinking mode)
Lines you could clip.
00:00
“I gave Claude Code access to a business of mine that does around a million dollars a year and I asked it to automate everything, and it actually did a really good job.”
02:33
“An eval is a test that repeatedly, over many, many times, records how well an AI model performs a specific task given a specific prompt.”
05:39
“All I now have to do in this business, to be honest, is just show up to calls and then talk about the systems I want to create.”
06:08
“I cannot tell you how freeing all of this is. I genuinely feel like for one of the first times in my life that I am operating at essentially maximum productivity.”
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.
So I gave Cloud Code access to a business of mine that does around a million dollars a year and I asked it to automate everything and it actually did a really good job. And this isn't a clickbait or one of those like hype beast videos where I just talk about how it's automated my whole life and everything's amazing. I've tried this many many times over the course of the last four years with different models all the way back to GPT 3 .5 Turbo and I'm usually quite disappointed but we've now reached the point where you can do this.
And so in this video, I'm going to run you through a very simple standardized process that you guys can use to automate most of your knowledge work as well. The first thing you need to do is you need to audit your work. This took me around a week and it involved me asking Opus 5 .5 on Max Thinking, by the way, to build a tracking tool that took a screenshot of what I was working on every five seconds or so throughout the day.
So about 12 times a minute. That's a lot of screenshots. So you have to have it compress it down to a minimum legible size and then tell you something about what you are looking at, what you are working on, and so on and so forth.
So I stored all this stuff in a database. I additionally asked Claude to ask me a bunch of questions throughout the day, about 10 or 15 of them. These are usually simple ones like, hey, what are you working on?
Why are you working on this? Why did you do that? What systems do you need in order to make this happen?
And the whole idea here is you can tell Claude what you're working on, but oftentimes you don't even fully know. all of the context required to make a decision. So it's far more effective to show it.
This took me around a full session's worth of usage. I had to upgrade my plans. You guys will probably have to do that as well if you want to do it on max thinking, which I'd recommend.
But assuming that you do, you will actually get a very, very in -depth breakdown of everything that you do. And after that, I had a big list of literally all of the deliverables and processes that I personally involved myself with in my agency. In my case, it was things like scraping lead lists using Appify or LinkedIn Sales Navigator.
It was enriching and verifying emails. It was writing cold email copy and sequences. It was buying and then warming up domains and then sending accounts, replying to cold email responses, booking calls, checking new bookings, running discovery calls, sales calls, proposals.
Obviously, there are a lot of things that go into running an agency, so I'm going to spare you all of that, but it's quite the list. The key thing is here, after... you have the list, just feed that into a new cloud instance and then say, hey, here's everything that I do on a daily basis.
My goal is automating the vast majority of my non -customer facing work. Can you help me build systems for each of these to automate as much as possible? And if you don't have enough context on any of this stuff, just ask me and then it'll ask you.
And then all you need to do is just voice transcribe your answers to it. So I just use a simple voice transcription tool. You can dump it all out in probably about an hour or so, and then it will have all of the context that it needs in order to do the automation.
So for each of these tasks, then you need to create what are called evals. An eval is a test that repeatedly over many, many times records how well an AI model performs a specific task given a specific prompt. It'll then average the results.
You need this because sometimes AI models will do really, really well in isolation. Other times it'll do a really, really poor thing or outcome on the same job. So in my case, for every task that I gave it, everything that I've just mentioned, I had to do it 10 times.
And then I looked at those 10 outputs and I scored them myself. If the score was less than approximately an 8 out of 10, sometimes I was fine with a little bit less certain or accurate score, but most times it was about 8 out of 10 or 80%. I just asked Claude, hey, how can we improve the system to increase this score according to my goals?
Hey, I've noticed that a lot of the outputs look like this, and I actually want them to look like that. So we just looped over and over and over again until it successively improved the skills and the systems that it used to produce the outputs. And then I got all of them, approximately all of them to 7 out of 10 or 8 out of 10 or above.
That was probably the most time intensive aspect of this whole process. The screenshots and stuff took me no additional time because it's just monitoring my work and looking over my shoulder. But this was all like active time where I was actually like, hey, this is wrong.
We need to fix this. Hey, that's wrong. We need to fix that.
Oftentimes, I also had to give it a bunch of templates or assets. So I had to give it a bunch of like high performing cold email copy examples. For instance, I had to give it a bunch of examples of like qualified lists, high quality proposals.
Like, you know, you're going to have to do a little bit of legwork here. But then for all of the systems that score over a certain amount, you can actually institutionalize these as skills. And the reason why it's so important to institutionalize these as skills is because if you get on something like a cloud teams plan, you can then share these skills with other people in your organization, even non -technical people.
that don't necessarily need to use Cloud Code in order to get value out of this. They could use simpler tools like Cloud Code or Cloud Chat. By the way, something that'll save you a lot of time when you do this is not actually just improving one tiny little part of the skill and then retesting the eval suite on 10.
I'd actually recommend you batch three or four tests and run them all simultaneously because that way you'll parallelize the changes that Cloud is making to the skills. And so this saves me a lot of time. If I had to go one at a time, it'd probably take me like a week as opposed to a day.
Finally, for every skill that I now had, I simply asked Claude, does it make sense to schedule this to run autonomously without me? Some skills like preparing a draft to all of my clients, which I do every morning. And it's a simple draft that just explains, hey, you know, here's what I'm working on.
Here's what I did yesterday. Here's some blockers that I have. Those are basically completely automatable.
You just, you know, give it a check through before sending and then press enter. But other ones like maybe building a deliverable of some kind, building a system. writing a cold email campaign from start to finish.
You don't actually want to schedule these automatically because they do still require at least human oversight or human in the loop. And so these skills I run when I need to, essentially. But yeah, you schedule all of the things that you can schedule.
And then at the end, you literally just have a short list of tasks, like in my case, about five or six tasks that I do on a daily basis for the agency. The other 80 or 90 are completely automated. And yeah, at the end of this process, I literally ended up automating virtually everything in the business.
that could run without me. And that's an important distinction. All I now have to do in this business, to be honest, is just show up to calls and then talk about the systems I want to create, which, you know, some people don't like doing, of course.
But for me, this is actually the part that gives me the most value, the most joy. It's connecting with people, figuring out their bottlenecks, which, by the way, you can consult Claude on as well, but actually chatting with people and like really figuring out what the heck's going on and then using my own human knowledge to proactively suggest, you know, solutions based off my experience and case studies.
I cannot tell you, as somebody that has had to do so much software engineering in the past, so much useless A -B testing, dragging and dropping of modules, how freeing all of this is. I genuinely feel like for one of the first times in my life that I am operating at essentially maximum productivity. I don't really know.
how much further you can go after this process. I'm sure technology will continue to surprise me. But yeah, I'm definitely feeling the AGI.
And if you guys run through this process top to bottom, just like I did, you will as well. So if you want the same process that I talked about in this video completely free, you can head over to Maker Zero. That's the second link in the description and just copy over a detailed prompt that will walk you through building this whole thing.
I would highly recommend to use the most frontier model you can on max thinking because this is such an upstream planning style skill. Everything downstream of it will essentially be impacted.
And so it makes the most sense to use the most capable model for the task. And that's even if you burn out of usage, you might just have to wait a little while and then run when your session resets. Aside from that, if you guys want to grow a business doing this exact thing, check out the first link in the description.
That's Maker School. That's my 90 -day accountability program where I'll guarantee you, you get a customer in 90 days or I'll give you your money back for any AI automation or AI consulting service, kind of like what we just talked about today. Feel free, take a look inside.
um thank you guys very much for supporting the channel as you always have and i'm looking forward to catching all y 'all in the next video see ya
The Hook
The bait, then the rug-pull.
Nick Saraev opens with a clarification before the pitch: this isn't one of the hype-beast AI videos, he's tried automating his work with every model since GPT-3.5 Turbo and been disappointed every time, but this week, with Claude Code, it actually worked.
Named ideas worth stealing.
00:00model
Audit-to-Automation Process
- Audit your work (screenshot + Q&A for a week)
- List every task
- Have Claude plan a system per task
- Build evals, score 10 runs, iterate to 7-8/10
- Turn passing systems into skills
- Batch-test improvements
- Schedule what can run unattended
A seven-stage loop for finding everything you do, designing an AI system for each task, proving it's reliable with repeated scored tests, then deciding what runs on its own.
Steal forAny Modern Creator ops process Joe wants to hand to Claude: the VOD pipeline, the email P.S. rotation, the mca_jobs admin queue
How they asked for the click.
VERBAL ASK
07:04product
“If you want the same process completely free, head to Maker Zero and copy a detailed prompt that walks you through building this whole thing. If you want to grow a business doing this exact thing, check out Maker School, a 90-day accountability program with a get-a-customer-or-your-money-back guarantee.”
Two-tier soft pitch: a free prompt pack for anyone, then a paid accountability community for people who want hands-on support, both framed as a direct continuation of the process just taught.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
