Write Your Own Skills. That’s the Edge.

There’s a video making the rounds right now (Ras Mic on the Greg Isenberg podcast) and it’s the clearest articulation I’ve seen of what separates people who use AI tools from people who are truly formidable with them.

The thesis is simple: the models are exceptionally good now. Opus 4.6, GPT 5.4... we’ve reached a point where the differentiator isn’t which model you pick. It’s the context and harness you build around it. And most people are building that harness wrong.

Here’s what I mean. When you create a claude.md or agent.md file, that entire document gets loaded into the context window on every single turn. A thousand-line file? That’s 7,000 tokens burned on every exchange, whether the agent needs that information or not. Multiply that across a conversation and you’re hemorrhaging context, pushing the agent closer to its limit where performance degrades. You’re literally making your agent dumber with every message.

Skills flip this on its head. A skill file stores a name, a description, and detailed instructions, but only the name and description enter the context window. The agent reads the full file only when it determines the skill is relevant. That’s progressive disclosure. That’s 53 tokens versus 944. That’s the difference between an agent that stays sharp for an entire session and one that starts hallucinating halfway through.

But here’s the part most people miss: how you create the skill matters more than the skill itself. You don’t sit down and write the perfect skill file from scratch. You don’t download someone else’s skill from a marketplace. You walk through the workflow with the agent, step by step, correcting it in real time. You let it fail. You show it what success looks like. And only after a complete successful run do you tell the agent: "Review what you just did. Now create the skill."

The agent writes the skill from lived context, not from your abstract description of what should happen. That’s the difference between a skill that works 60% of the time and one that holds up run after run.

Then you iterate. The skill will still hit edge cases. When it does, you identify the failure, have the agent fix it, and tell it to update the skill file so it doesn’t happen again. Five iterations of this loop and you have something that no downloaded template will ever match: a skill forged from your specific workflow, your specific data sources, your specific definition of success.

This is how I build. One agent, one workflow at a time. No fifteen sub-agents on day one. No marketplace installs. Just methodical, recursive refinement until each skill is reliable... then I scale. The result is a system that’s more productive than any shiny multi-agent architecture, because it was built from the ground up on actual successful runs, not theoretical configurations.

Scale for productivity. Not for what looks cool.