AI Agents & Claude Skills Full Course: Setup and Build
TL;DR
The models are already good enough; context is now the real lever — Ross Mike says tools like Opus 4.6 and GPT 5.4 are “exceptionally good,” so the difference between great output and slop is mostly how you assemble context, not which frontier model you picked.
Most people should stop stuffing agent.md files with boilerplate — He argues 95% of users don’t need an agent.md/claude.md file because repeating things like “this codebase uses React” wastes tokens the model can infer directly from the repo.
Skills beat giant prompts because they use progressive disclosure — A skill only adds its name and description to context until needed, so a 944-token instruction file can become a 53-token hint that the agent expands only when relevant.
The right way to build skills is to teach the workflow first, then codify it — Ross’s sponsor-screening agent only got good after he walked it step by step through checking Twitter, YouTube, Trustpilot, funding, and rejection criteria, then had the AI turn that successful run into a skill.
Expect 2 weeks of frustration before agents become truly useful — He says early agent use feels like “what is this garbage,” but once you iterate through failures, ask why it broke, and recursively update the skill, reliability jumps fast.
Scale for productivity, not for what looks cool on Twitter — Instead of starting with 15 sub-agents and 30 downloaded skills, Ross recommends one main agent, a few proven workflows, then sub-agents only after you’ve earned the complexity.
The Breakdown
The big claim: the models are finally good
Ross opens with the heretical take: most of the discourse around agents is wrong because the core models are already strong enough. He name-checks Opus 4.6 and GPT 5.4 as “amazing,” then reframes the whole problem: you’re no longer fighting model quality so much as steering quality through context.
Why giant agent files are usually a waste
He breaks context into parts: the provider’s system prompt, your agent.md/claude.md file, skills, tools, codebase, and chat history. His blunt view: 95% of people do not need a giant agent file unless they have proprietary company info or a methodology that truly must appear in every single run.
Skills are the real unlock because they stay lightweight
Ross is openly a “skills maxi,” and the reason is token efficiency. A skill only exposes its name and description at first, then pulls in the full instructions when the agent realizes it needs them — what he calls progressive disclosure — instead of paying thousands of tokens on every turn.
The sponsor-email story that changed his workflow
His example is great: he built an OpenClaw agent with its own email to review sponsor inquiries every 15 minutes. At first, the agent said every company looked legit, so he had to manually teach it the actual workflow: check Twitter, YouTube, Trustpilot, funding history, and reject automatically if key signals are missing.
Don’t write the skill first — train the agent through a successful run
This is the core method. Most people identify a workflow and jump straight to writing a skill, but Ross says that’s backward; instead, you walk the agent through the task like a new employee, let it fail, correct it, repeat, and only then ask it to generate the skill from the successful context.
Why downloaded skills and flashy agent stacks miss the point
He’s skeptical of skill marketplaces and “beefed up” setups full of sub-agents because they look impressive without being grounded in your actual work. His line is memorable: build for productivity, not for what looks cool — one agent first, then real workflows, then sub-agents for marketing, business, or personal tasks if they’ve earned their place.
Recursive skill-building: use mistakes as training data
Even after a skill exists, it will still break, and Ross says that’s not failure — that’s the moment to improve it. His process is to ask the agent why it failed, feed the error back, have it fix the issue, then update the skill so the same bug doesn’t happen again; he says his YouTube reporting workflow took five loops but now pulls from eight data sources in about 10 minutes reliably.
Less is more, especially as the context window fills up
He closes by arguing that harness quality and context design now matter as much as the model itself, citing differences between Cursor, Claude Code, and Codex outputs. The practical takeaway is simple: don’t waste tokens telling the model obvious things like “use React”; save context for what’s unique to you — your workflow, taste, business logic, and decision rules — because once the context window gets crowded, the agent starts getting “dumb.”