How AI agents & Claude skills work (Clearly Explained)
TL;DR
The models are already good; the bottleneck is context — Ross says Claude Opus 4.6 and GPT-5.4 are strong enough that the difference between great output and “slop” now comes mostly from how you assemble context, tools, and workflow instructions.
Most people do not need an agent.md/claude.md file — His rule of thumb is 95% of users should skip always-on instruction files unless they contain proprietary company knowledge or a methodology the model truly needs on every single turn.
Skills beat giant prompt files because they use progressive disclosure — A skill only adds its name and description to context until the agent decides it needs it, which can mean 53 tokens instead of 944 or even 7,000+ tokens being loaded every run.
The right way to build a skill is to teach the agent live first — Instead of jumping straight from “I have a workflow” to “I wrote a skill,” Ross walks the model through the task step by step, gets a successful run, then asks the AI to turn that exact process into a skill.
Expect agents to fail early, then use those failures as training data — Ross says his best workflows came from recursive iteration: ask why it failed, fix the error, then update the skill so the same mistake doesn’t happen again, which is how he got a multi-source YouTube reporting agent working flawlessly after five loops.
Scale for productivity, not for what looks cool — His advice is to start with one agent, add your own workflows and skills, then introduce sub-agents only when they map to real jobs, instead of spinning up 15 sub-agents and 30 skills before you’ve defined how you actually work.
The Breakdown
The big thesis: models are strong now, but context decides the result
Ross opens by saying the debate has shifted: the models are no longer the weak point. He names Claude Opus 4.6 and GPT-5.4 as “amazing,” and says we’ve hit a stage where you can steer the same model toward either quality or total slop depending on the context you give it.
Why he thinks most people are overusing agent files
He breaks down the context stack: system prompt, agent.md or claude.md, skills, tools, codebase, and the live user conversation. His spicy claim is that 95% of people don’t need a giant agent file at all, because if the codebase already shows it’s React, telling the model “this codebase uses React” is like reminding a podcaster to bring a microphone.
Skills are better because they don’t bloat every run
This is where he gets evangelical: Ross calls himself a “skills maxi.” The key mechanic is progressive disclosure — only the skill’s title and description sit in context until the agent decides it needs the full instructions, which saves huge token costs and keeps the context window cleaner.
His sponsor-email workflow shows why vague prompts fail
Ross tells a very human example from his YouTube business: he forwards sponsor emails to an OpenClaude agent and originally told it, basically, “research this sponsor and tell me if they’re worth it.” The result was comical — every sponsor came back “legit” and “perfect” — which forced him to realize the agent needed a concrete checklist like checking Twitter, YouTube, Trustpilot, and fundraising history before making a judgment.
Don’t write a skill first; mentor the agent like a new employee
His core workflow is: identify the job, manually guide the agent through it, then convert the successful run into a skill. He says people skip the experiential-learning step, but that’s exactly what gives the model the context of what “good” actually looks like — the same way you wouldn’t hire a new employee and say, “Good luck forever.”
Recursive skill-building is how brittle workflows become reliable
Even after a skill exists, Ross expects it to break. His move is to ask the agent why it failed, feed the failure back in, have it fix the issue, and then update the skill so the mistake doesn’t repeat; he says his YouTube report generator now flawlessly pulls from Notion, Dub, YouTube, Twitter, and other sources because he did five rounds of this refinement.
Start with one agent, then earn the right to use sub-agents
Ross pushes back on the trend of flashy setups with 15 sub-agents and 30 downloaded skills. His view is simple: begin with one useful agent, build your actual workflows, then add sub-agents for real responsibilities like marketing or business only after the core system already works.
Less is more, especially as the context window fills up
Near the end, he zooms back out: harness quality and tool design matter more and more now that the underlying models are strong. He gives a concrete example from his own code-structure skill — 116 lines long, 944 tokens if always loaded, but only 53 tokens when represented by name and description — and argues that keeping context lean makes agents cheaper and smarter because they get “dumb” as the window approaches full.
A personal close about impact over hype
The episode ends on a surprisingly grounded note: Ross almost didn’t come on because he didn’t have a flashy new tool to discuss. Greg reads him a message from a viewer who used one of their earlier videos to get into coding and build a cake business doing $150,000 a year, which reframes the whole conversation from chasing 200,000 views to making something genuinely useful for the right people.