Skills, Ambient Intelligence, and the Folder That Thinks

AI Mastermind | Knowledge Entrepreneurs Edition

“Your irreplaceable edge — your expertise, your judgment, your proprietary way of thinking — that transfers, too. But not automatically. Not by accident. You have to engineer it, deliberately. That’s what skills are.” — Lou

This Week in 30 Seconds

Skills, demystified — a skill is just a folder with a SKILL.md file. Everything else is optional power-ups.
The Brand Writing Team, built live — Lou stood up a 7-role writing skill from first principles, injecting quality gates and “outside-the-modal” thinking mid-build.
The Living Knowledge Base vault — a Karpathy-inspired schema that ingests transcripts and generates insights, commands, skills, article briefs, and a Voice-of-Customer library automatically.
The audit-and-codify move — the single most reusable habit Lou demonstrated: after any long back-and-forth, tell Claude to scan the conversation and bake every fix and decision back into the skill, doc, or code at root cause.
Don’s onboarding assessment skill — 20 participants, MBTI + OCEAN + a custom Career Climbers Index + a 15-minute interview, all fed into a Claude-built coaching-insights pipeline that spotted cross-participant signals no human reviewer would have caught.

What a Skill Actually Is

A skill is a folder on your hard drive with a SKILL.md file inside. That’s the whole trick. Everything else — resources, templates, scripts, sub-prompts — is scaffolding you add only when the thing gets big enough to warrant it.

The elegance of this is that it inherits from the oldest idea in Unix: move text around, let the filesystem be the database, and trust small composable pieces. Lou’s framing was explicit — “I think it’s borrowing a little bit from the Linux concept of, hey, let’s just move text around and we’ll figure things out.” If you’ve been intimidated by “building an agent,” the reframe is freeing: you’re writing a markdown file and putting it in a folder.

Progressive disclosure is the feature that makes this scale. Only the name and description of a skill get loaded into context on startup. When Claude decides it needs the skill, it pulls in the rest of SKILL.md. When SKILL.md decides it needs a specific role (researcher, editor, brand voice) it pulls in that file. Nothing enters context until it earns its seat. This is why the right mental model for SKILL.md is orchestrator, not worker — it’s the thing that decides which subordinate files to activate, and when. Lou’s rule of thumb: if SKILL.md creeps past 200–500 lines, start breaking pieces out into resources/, templates/, or scripts/.

💡 What This Means for You

If you’ve been putting off “making a skill” because it sounds like engineering, stop. Open a folder on your desktop, create a SKILL.md, and write the same prompt you’d write in a chat — but with a name and description at the top. That’s a working skill. You can add the rest when you actually need it. The gap between “I have a good prompt” and “I have a reusable skill” is five minutes and a directory.

The Audit-and-Codify Move

At the end of any build, instead of accepting the skill as-is, tell Claude: “Audit our conversation. Every major decision we made, every error we fixed, every extra requirement I added along the way — go find where those belong in the skill, and codify them at the root-cause level.” This is the difference between a one-shot fix and a durable improvement. Most people debug a skill, get it working for the moment, and never close the loop back to the source-of-truth file. The audit move makes the conversation itself into training data for the skill.

🔥 Hot Take: Your Skills Are Underperforming

Most people treat skill-building as a one-pass exercise: write the prompt, ship it, use it. Lou’s build added three distinct pressure mechanisms — orthogonal angles, quality rubrics with forced revisions, and a skeptic role whose entire job is to poke holes. The 9/10 output wasn’t because the underlying model is better. It’s because the skill was engineered to refuse its own first answer. If your skills don’t have an adversarial step built in, you’re reading first drafts and calling them finished.

Recap produced with Claude Code · AIMM Mastermind · April 9, 2026