Apr 8, 2026 · 35m · startup-ideas
How AI agents & Claude skills work (Clearly Explained)
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
AI practitioner Ross Mike joins Greg Isenberg to break down how AI agents and Claude skills work, offering actionable strategies for context optimization, experiential skill development, and pragmatic agent architecture.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Greg holds 13.5% of the talking time here. How this is scored →
Greg as informed peer 3.4 Guest teaching 5.0 Guest disagreement 2.0 Greg pushing back 1.4
speaking balance: gold is Greg, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 32:30 Denouncing agent configuration files as a farce
Ross strongly rejects industry conventions, calling popular agent.md and claw.md files a farce that users should strip out in favor of clean skills.
Hardest push from Greg ▶ 17:49 Correcting the vibe coding app valuationGreg immediately cuts across Ross's claim of a 100-million-dollar vibe-coded company to firmly correct the figure to 1.8 billion dollars.
Biggest teaching moment ▶ 30:45 Mathematical breakdown of progressive skill disclosureRoss uses a tokenizer to mathematically prove that progressive disclosure reduces token consumption from 944 tokens to 53 tokens per interaction.
Greg holds their own ▶ 24:39 Deploying the Office rundown context metaphorGreg connects technical context engineering to a relatable pop culture reference from The Office, demonstrating intuitive grasp of context starvation.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Greg as informed peer | Guest teaching | Guest disagreement | Greg pushing back | Why |
|---|---|---|---|---|---|---|
| Foundation of Modern Models and Context Windows | 3 | 6 | 2 | 1 | Ross asserts that 95% of users do not need agent.md files because modern frontier models possess strong base capabilities. Greg asks a clarifying question to uncover what specific situations fall into the remaining 5%. | |
| How Progressive Disclosure Works in Agent Skills | 0 | 5 | 1 | 0 | Ross delivers a technical monologue breaking down how progressive disclosure injects only skill titles and descriptions into context until needed. Greg remains silent while Ross diagrams context window management. | |
| Developing Skills Through Step-by-Step Experiential Learning | 3 | 6 | 2 | 1 | Ross describes training his OpenClaw email filtering agent iteratively before codifying the steps into a skill. Greg validates the premise by sharing how frustrating it is when AI requires handholding on seemingly binary tasks. | |
| Scaling for Productivity Over Aesthetic Complexity | 4 | 4 | 3 | 2 | Greg synthesizes Ross's advice into an employee onboarding comparison and asks directly if he is telling builders to do unglamorous work. Ross criticizes vanity agent setups with 15 sub-agents in favor of gradual, productivity-first scaling. | |
| The Changing Economics of Software and Vibe Coding | 5 | 2 | 1 | 3 | Greg interrupts Ross to correct his market valuation estimate on a vibe-coded company from $100 million to $1.8 billion. Greg also pushes back on dramatic industry discourse around a permanent AI underclass. | |
| Code as Context and Recursive Skill Refinement | 3 | 6 | 2 | 1 | Ross explains recursive skill building, detailing how error feedback should update markdown skill definitions rather than causing user frustration. Greg reframes this as a necessary shift in user expectations from instant magic to iterative debugging. | |
| Managing Early Friction and Sub-Agent Architecture | 5 | 4 | 2 | 1 | Greg introduces an analogy from The Office where an employee is tasked with a rundown without context to mirror agent confusion. Ross reinforces the comparison by explaining how he grew from one foundational agent to five specialized sub-agents. | |
| Token Economics, Context Degradation, and Markdown Walkthrough | 4 | 7 | 3 | 2 | Ross demonstrates token math comparing 944-token system prompts to 53-token progressive skill calls, warning of model degradation past 70% context capacity. Greg highlights that preserving context is crucial for reasoning quality in addition to cost savings. |
Statements from this episode (10)
Ross Mike: 95% of AI Agent Users Do Not Need Agent.md Files
“And then you have a lot of people have agent.md files or cloud.md files. Now I'm just going to say off rip, 95% of people don't need this. The reason being is, again, you have to assume that the models are already good, right?”
Ross Mike: Agent skills load only metadata initially to conserve tokens
“What, when you create a skill.md file, what gets added into the context is actually just the name and the description. Right? The bunch of info doesn't get added. So imagine you have two sentences versus an agent.md that has like a thousand lines that get adde…”
Ross Mike: Do not download AI agent skills from public marketplaces
“I don't download skills because your agent needs the context of a successful run, which you then turn to skills, right? And this is the big thing I see. You see skills marketplaces. You see download this and that. First of all, it's a easy way to attack somebo…”
Ross Mike: The best AI skills come from guided successful runs
“To me, the best way to create a skill is to work with it in your specific workflow. Once you have a successful run, tell it, okay, review what you just did. This is the skill you need to create.”
Ross Mike: Treat AI models like new employees, not all-knowing magic
“We should treat models and these agents like very new employees versus like these black magic boxes that like know everything. Right? They know everything because they've been trained on a lot of data, but they don't know your workflow, your steps, right?”
Isenberg: Calling AI-displaced workers a permanent underclass is ridiculous
“It's ridiculous to call it a permanent underclass.”
Ross Mike: Software starter templates will have a renaissance due to AI agents
“I believe templates are going to have a renaissance because if you have a solid like template, right, like whether it be like for a web app or mobile app, because that becomes context for the agent, it's going to build on top of that, right?”
Ross Mike: Agent Frameworks Hide Painful Two-Week Initial Setup Friction
“So there's like this early area of investment that you have to make that sucks that nobody will tell you, especially agent harnesses company because they wouldn't raise as much money if they did. But like this, maybe I would give it two weeks. Because it took …”
Ross Mike: Tooling and harness will matter more than base model improvements
“The next iteration is probably going to get better, but the harness and the tools that you surround it, the context that you give it is going to matter even more.”
Ross Mike: LLMs degrade in performance once context fills past 70-80%
“Cause the model will get dumb as the context window closes, right? So if you have like a context window and I can draw this out, if this is your context window and like the optimal is You're between, like, there's always, like, maybe, like, 10% is already fill…”