Most people treat AI like a chat buddy. Type a prompt, wait, type again, repeat forever. That approach works until you realize the real cost isn’t building the agent, it’s babysitting it as tools break and models change.
Here’s the contrast that stuck with me. This walkthrough comes from Grace Leung, a digital growth consultant who tests AI marketing setups for a living. Instead of building an agent from scratch, the creator flips the whole idea: what if the team already exists and your only job is to direct it? I was genuinely impressed by how much cleaner that framing makes everything.
She breaks it into three levels of leverage, demoed inside Ahrefs’ new marketing agent (called Agent A, now Letaido). But the levels aren’t tied to that tool. You can take all three into Claude or whatever you already run.
The old way vs the new way
The old way: one endless chat thread where you re-explain context every single time, engineer workflows live, and hope the output stays consistent. You are the bottleneck.
The new way: treat your agent like a team with defined jobs. Group work by marketing function, hand each function a repeatable process, and let the system carry the context. You stop being the operator and start being the director.
That shift is the whole point. Let’s walk the three levels the author lays out.
🎯 Level 1: Leverage agent skills
A skill is just a proven workflow you hand to the agent so it runs on command. The original poster calls this the fastest ROI, and it makes sense: you’re not building, you’re using something that already works.
Her starting move is smart. Before touching any skills, she asks the agent to map her recurring marketing jobs and organize them into a skill library grouped by discipline. That gives you a clear structure instead of a random pile of prompts.
Then she runs a pre-built one: an AI search visibility gap analysis. Type in a brand, drop the target domain, pick the AI platform to check against, and it returns a prioritized mention-gap report plus a dashboard you can sort by question, competitor, and platform. The creator points out you can even set it to auto-run monthly.
One honest note she shares: running it on Claude Opus got pricey. Her tip is to switch to a lighter model like Sonnet for most jobs and save the heavy model for when you actually need it.
The deeper move at this level is chaining. She uploads her own custom “brand strategy deck” skill, saves it to the knowledge base, then asks the agent to run a full growth analysis and package it into a 14-slide branded deck. Same idea you’d use in Claude with saved skills.
Level 2: Define agent roles
Here’s a subtle point the expert nails. Skills are just actions. On their own, they don’t know when to fire. You still have to decide which skill to use and when.
So Level 2 is about creating a focused worker with a role. When she defines an agent, she includes six things worth copying:
- Role: what this agent is responsible for
- Inputs: what it needs from you
- Skills/tools: what it’s allowed to call
- Workflow: the steps it follows
- Output: the exact format you want back
- Guardrails: the limits it stays inside
She builds a “content strategist” agent, saves it as a playbook, and updates the agent’s memory so it loads that playbook whenever she tags it. Then she spins up a separate “blog writer” agent that picks up the priority topics, researches with real data, calls a WordPress connector, and drafts an article that follows her rules: bullet takeaways, question headings, tables, internal links, and an FAQ.
Two agents, two clear jobs, one shared workspace. That’s the leverage. The creator still stresses reviewing every draft with your own judgment before publishing, which I appreciated.
Level 3: Build tools and apps
Level 3 is where your best workflows leave the chat window entirely. You package the logic into a simple tool anyone on the team can run.
The author installs a content keyword research app, types a niche plus seed keywords, and gets back clusters by traffic potential, competitor landscape, search intent, recommended content formats, and a hub-and-spoke topic map she can export as a client-ready PDF.
Then she goes further and asks the agent to build a custom tool: research real community questions, wait for her approval on an angle, then generate carousel social visuals. About ten minutes later, the tool exists. Now a teammate can run her exact process and get her level of output without knowing any of the setup.
Buy vs build, said honestly
I liked that the mind behind this video didn’t oversell. Her take: if you’re technical, solo, or want full control, build it yourself in Claude with an MCP or data API and own every piece. If you’re on a team and don’t want to maintain infrastructure as models shift, a hosted, shareable setup gets you working today.
Either way, the three levels don’t change: leverage skills, define roles, systemize into tools. That’s the ladder from prompting to actually directing AI.
Want to see the dashboards, the deck, and the custom tool getting built in real time? Watch the full video. It’s the clearest demo of this three-level system I’ve come across.