# How to Master AI in 30 Days (as a non-tech)
**作者**: Mark Vassilevskiy
**日期**: 2026-03-08T17:04:39.000Z
**来源**: [https://x.com/MarkKnd/status/2045183177132785941](https://x.com/MarkKnd/status/2045183177132785941)
---
I'm a designer and built a 7-figure launch video agency. I’ve never had any prior experience with coding, Cursor, Claude or anything else until 2026 came in.
Everyday I was seeing some new crazy features that companies like Anthropic are releasing a new update that gets 20m+ views, and all people around the world are saying that it just killed 1,000+ startups. The FOMO reached its peak.
But nobody says where to start or how to actually use them in your workflow
Out of everything I tried over ~3 months, maybe just a few tools moved me forward. The rest I could've skipped entirely.
So I reverse-engineered my own path into a 30-day roadmap. That’s the plan from someone who hasn’t written a single line of code, but at the same time automated my full agency using it.
Here's my entire playbook:
## 1. Stealing SKILLS files
Every "AI product" you pay $30/month for is usually a basic wrapper.
Notion AI, Jasper, Copy.ai, Viktor, all of them. Under the hood they're sending prompts to the same models you have access to (Claude, GPT, Gemini, etc.)
The difference is that they've written really good instructions.
Those instructions are called Skills.
A SKILL.md file. Literally a text document that tells AI how to behave, what to output, what format to use, and what to avoid.
And here's where it gets fun - you can reverse-engineer them.
Find the coolest AI wrappers in your field
Sign up for free trials (most have 7-14 days)
USE them in your real workflow
Notice what makes their output better than raw ChatGPT
Then you can just ask the assistant to disclose its own SKILLS file for you
In most cases, it’ll actually share it with you
Just download it and paste it into your Claude Code to create a 1:1 agent
PS. I highly recommend using only Claude Code and forgetting about anything else. Get the Max subscription, which is a bit pricey, but the ROI is insane, trust me.
## 2. Creating SOPs right
SOP - Standard Operating Procedure
Basically, a step-by-step plan that you would give to anyone who joins your company. Imagine that your new Claude Code workspace is your VA. How would you explain the whole pipeline? (where to start, which apps to use, providing all the necessary MCPs, etc.)
If you don't know the exact SOP, then you can just use the plan mode in Claude and it'll help you to create the final roadmap.
Here's an example of an SOP I was recently creating for my Attio automation:
- I get the call booking (a new lead appears in the “Leads” tab in Attio)
- After I sent the proposal, it should automatically move to the “Proposals” tab
- When the client hasn’t responded for more than 3-4 days, you should send a reminder in Slack (you should text me directly in Slack and draft a reminder message, I’ll then tell you if we should continue with the one you created, skip it, or if I wanna provide my own version of the reminder)
- Then we should send them a reminder every week
- If the lead hasn’t responded for more than 2 weeks, we move him to the “Hold” tab
- You should analyze the whole email of mark@skale.solutions and check our conversations with the clients (usually email threads are Skale <> [Client Name]). If you see that they agreed to proceed, made a payment (on Contra or Ramp), or signed the contract, then we put them into the “Won” tab.
- If you see that they said that they went with another agency, don’t need our services right now or just declined working together - we move them to the “Lost” tab
- It needs to be updated every morning at 9 am Kazakhstan time
## 3. Build a habit of vibe coding
I literally can't stop vibecoding.
I think Naval recently said that vibecoding in 2026 can bring you more joy than video games if you know what you're building.
For me, it's equal. I recently played Counter-Strike 2 with my friends and between rounds I was typing prompts in Claude.
In a nutshell, you should always take your Mac with you and whenever you get a second, just type a prompt so Claude keeps working. While you're washing the dishes, you're thinking about the architecture, then give the prompt and continue doing your thing while it works
It's lowkey perfect for people with ADHD. You're constantly in a state of multitasking and trying to get the most out of every second. I don't think it's good for mental health in the long run tho.
## 4. Track what you can automate & create agents
The next step would be to actually identify what you can automate.
AI Agents are incredibly good at helping you to code, manage, handle the bookkeeping & financials, and so on. You should identify all the repetitive tasks you don't wanna spend a single second on.
For example, I completely automated our finances in the agency. Claude analyzes my email once per day. Whenever it sees an invoice from Contra it automatically ads the value into the finance sheet in Google Sheets.
I've built a tracking system inside of Notion that shows: how many leads we got, how much money made, what's the profit margin, how many active projects and so on. I spent maybe like 4-5 evenings with my team to fully automate all of that and now it saves us thousands of dollars a month on the project managers and hours of work
So the roadmap:
- Find what can be automated (Try to automate as much as you can, as it also teaches you a lot about your business and how to use Claude)
- Build all the infrastructure on top of Claude Code
- Deploy your autonomous agents on the web server
- Spent a couple of weeks perfecting the system
- Enjoy tens (if not hundreds) of free time that you can allocate to more important stuff
## 5. Setting up OpenClaw (optional)
OpenClaw is basically a tool that brings your usual Claude Code on a new level of abstraction.
It's basically an AI agent that runs locally on your system (rather than the web-based approach I mentioned earlier), and it's more autonomous.
You can just give it a direction, and it'll work on it until it figures it out.
For some people it's beneficial, some it won't make a difference.
I personally bought a used MacBook just to test and see how it works. In reality, it might be slightly better than Claude, but it also takes much more effort to set up for a non-tech guy.
So, I'd recommend firstly working only with Claude, but after you feel like you need a bigger autonomy, you can try out OpenClaw.
6. Read @bcherny and practice
Boris Cherny is the creator and Head of Claude Code at Anthropic.
If you're looking to become a pro at vibe coding, his tweets are a must-read.
7. Set up workflows
First of all, check out GSTACK - Garry Tan's open-source Claude Code toolkit. Three commands worth knowing:
/autoplan - you give a direction, it breaks the full task into steps and builds a plan before writing
/qa - runs automated tests on whatever you just shipped
/browse - Claude opens a browser, navigates sites, scrapes, fills forms, does research
Superpowers - forces Claude to think before it codes
By default, Claude immediately starts writing, while the superpowers approach is:
/brainstorm - clarifying questions, alternatives, design doc saved. You stop building the wrong thing
/write-plan - exact steps, exact file paths
/execute-plan - subagents per task, fresh context each, auto-reviewed before moving on
Second workflow is GSD (Get Shit Done) - Claude gets dumb when context fills up
Past 70% it forgets decisions, loses the architecture, starts hallucinating fixes
GSD pushes heavy work into isolated subagents while the main thread stays clean
Cheatsheet
Max plan = stop counting credits
MCP - connectors that let Claude access external tools and data sources directly (Notion, Slack, GitHub, your DB)
CLI - the terminal interface where Claude Code runs
Auto-accept permissions - Just turn it on, it's good and saves you hours
Plan Mode - Claude thinks and maps the full approach before doing anything. Use it on anything complex
Sonnet - Opus is just better, forget about Sonnet
Opus - the GOAT
Best combo - Just combine any workflow you found useful for your project, use plan mode for every complex logic, and always use opus max mode
## 相关链接
- [Mark Vassilevskiy](https://x.com/MarkKnd)
- [@MarkKnd](https://x.com/MarkKnd)
- [14K](https://x.com/MarkKnd/status/2045183177132785941/analytics)
- [Copy.ai](http://copy.ai/)
- [Mar 9](https://x.com/MarkKnd/status/2030691176047935870)
- [93K](https://x.com/MarkKnd/status/2030691176047935870/analytics)
- [Apr 12](https://x.com/everestchris6/status/2043061132911677924)
- [1.2M](https://x.com/everestchris6/status/2043061132911677924/analytics)
- [@bcherny](https://x.com/@bcherny)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [12:50 AM · Apr 18, 2026](https://x.com/MarkKnd/status/2045183177132785941)
- [14.5K Views](https://x.com/MarkKnd/status/2045183177132785941/analytics)
- [View quotes](https://x.com/MarkKnd/status/2045183177132785941/quotes)
---
*导出时间: 2026/4/18 13:52:55*