You don’t need to write a single line of code to build a real AI agent anymore. I don’t mean another chatbot that sits there waiting for your next question. I mean an agent that takes a whole job off your plate. I just watched a walkthrough from Grace Leung, a digital growth consultant, and in it the creator builds a content agent inside the Claude desktop app. It turns her videos and newsletters into animated Instagram carousels.
What got me is how plain the whole process is. There’s no terminal and no scripts. It’s just folders, plain English, and a few smart habits. If you’ve been using AI every day and still doing most of the work yourself, this one’s for you.
🚀 The Big Idea: Five Parts, Five Levels
The author starts with a simple definition. An agent is an agent because it uses judgment. You give it a goal, and it decides what to do first, which tools to use, and whether the result is good enough.
To make good calls, every agent needs the same five parts:
- The brain: the AI model doing the thinking. She picks Opus 5.5 because it follows detailed instructions well and has strong design taste.
- The context: what it knows about your brand, your audience, and your voice.
- The job: the goal it’s working toward.
- The rules: how it should work and what it should never do.
- The tools: what it’s allowed to use to get the job done.
From there, the expert builds the agent in levels. Level 0 sets up the workspace. Level 1 builds the agent. Level 2 makes it reliable. Level 3 makes it self-improving. Level 4 puts it in the cloud so it runs on its own. Each level stacks on the last one, and you can stop at any point and still have something useful.
💡 What Really Makes This Work
- 🧱 The workspace matters more than the prompt. Most people skip Level 0, and the creator thinks that’s the biggest mistake. Think of it like onboarding a new teammate: they need a desk, the right files, and a guide. In practice, that’s one folder (she calls hers “Business OS”) with her logo, photos, and a few real Instagram posts as references. She then asks Claude to read her website and those files and create two things. The first is a CLAUDE.md file, the onboarding guide Claude reads at the start of every session. The second is a brand context file covering her business, audience, voice, and visual style. Claude even asks follow-up questions, like which fonts she uses and why she likes certain posts. The proof is in the side-by-side test. The same request in an empty folder gives generic output, while the set-up workspace gives an on-brand weekly content plan with her real color palette. The point is simple: if you skip context, you get AI-looking content.
- 🔁 Skills make it consistent, and memory makes it smarter. At Level 1, the agent did a good job but made its own calls every run. So at Level 2, the author adds a “skill,” which works like a training manual for how the job should be done every time. Her trick is to let Claude interview her about what to improve and to tell it to “think hard.” The questions got noticeably better, like what signature emotion her posts should carry. The resulting content playbook skill (with motion built in) animated every slide and exported each one as a separate video, ready to post. Level 3 fixes the biggest weakness of AI agents: every session is day one. She adds a memory file the agent reads before every job, with one key rule. It should only save feedback that applies to future posts, not one-off edits. After she flagged an off-brand brown color and repeated slide backgrounds, the next carousel on a totally different newsletter had already fixed both. The challenge is that the first versions are never perfect. Images didn’t load, fonts looked basic, and a glass jar animation felt off-style. But every round of feedback compounds.
- ☁️ The cloud turns a helper into a worker. Level 4 is where this gets really fun. Right now everything lives on the laptop, so closing it means the agent stops. The innovator connects Notion as a tool (through Customize, then Connectors) and syncs the Business OS folder to a private GitHub repository. Don’t let GitHub scare you. It’s just an online copy of your folder, and Claude walks you through logging in. Now she drops ideas into a Notion “content inbox” and triggers the agent from the Claude mobile app, even with the phone screen off. Then she goes one step further with a scheduled task called “Daily Content Drop” that runs every evening. It picks up new inbox items, builds carousels, marks tasks done, and even suggests improvements to its own playbook. Whatever she drops in during the day is waiting as a draft the next morning. One practical note: she says she’s on the Max plan, so keep usage costs in mind if you plan to run daily cloud jobs.
✍️ Prompt of the Day: The Agent Designer
You don’t have to write the agent file yourself. The creator shares a ready-made “Agent Designer” prompt in the video description. You paste it in, describe the job, and it interviews you. A few parts of it stood out to me:
- It reads CLAUDE.md first and never asks what it can find there.
- It sticks to one clear job. If you describe several, it suggests which one to start with.
- It asks at most 3 questions, one at a time, about judgment calls, what “done” looks like as a yes/no checklist, and when to stop and ask you.
- It shows a one-screen plan covering Context, Brain, Job, Rules, and Tools before building anything.
- It keeps the agent under 50 lines and brand-agnostic, with built-in guardrails like “never invent facts, numbers, quotes or dates” and a “Look, Do, Check, Fix” loop capped at 2 rounds.
To use the finished agent, type “@” and pick it. To run a skill, type “/” and search for it. Then update CLAUDE.md so Claude hands social post requests to the agent automatically.
Try It This Week
You don’t need all five levels on day one. Set up your workspace, build one agent with a clear job, then add skills, memory, and the cloud as you go. The same steps work for any agent, not just content.
Watch the full video to see the carousels come to life and to grab the Agent Designer prompt. Then pick your level and start building!