Extraction
Record. Transcribe. Create.
One-hour Zoom of someone solving a problem. Give Claude the transcript. Have it create a skill. The recipe moves from one person's head into a reusable AI-operable system.
2026-05-14 / Yesterday I Learned
The Discipline Was Designed by a Human
Nashville AI Week wasn't about frontier models. It was a room full of people who get stuff done, and the pattern across every talk was the same: AI succeeds when humans extract tacit knowledge, write down decisions, build feedback loops, and use agents inside clear structures.
Every Team Has Someone Who Just Knows the Recipe
Chris Thomas put it cleanly at Nashville AI Week: every team has someone who just knows how to do something. The recipe lives in their head, not in a document. Your job is to get it out of there and into a reusable skill. Teams that lose are prompting. Teams that win are using skills.
Failure Modes
What Kills a Skill
Vague descriptions. Overlapping names. What without why. A skill fails when the context that makes it safe to run is missing from the spec itself.
Prototype
Unknown Unknowns
Matt Pocock's /prototype skill is for moments you can only figure out by building. It generates radically different variations and keeps a human in the loop to apply taste.
AI Isn't Fast. It's Disciplined. And That Discipline Was Designed by a Human.
Brian Moyer spent his lunch slot at Nashville AI Week giving a professional review of building concurrent.app, a medical family-sharing app, entirely with Claude and Replit. The collaboration loop: spec to implementation, with agents checking each other in cycles.
AI is good at: specs, scaffolding, migrations, version logging, using the knowledge base.
AI is not good at: column name differences, one registration bug, product-vs-code decisions, quality judgments. Those are human-driven, and they always will be.
Teams that lose are prompting. Teams that win are using skills.
ADRs, But for Everything
Tod Fetherling opened with a staggering frame: $400 billion lost annually from poor decisions in the US economy. Not from bad data. From decisions made without the right knowledge structure, "guesswork dressed up as analysis."
"A good decision is based on knowledge, not on numbers."
Decision Sciences Research frames effective decision-making around three distinct forms of understanding: Episteme (knowing that), Techne (knowing how), and Phronesis (knowing when and why). Data science covers the first two. The third is the missing layer, and it's the one that keeps breaking.
The framing that hit hardest: What happened? What do we do? Decision record. Architects have ADRs. We need the equivalent for product decisions, marketing calls, AI workflow changes, and cross-functional tradeoffs. Not to slow things down, to make the invisible knowledge visible and reusable.
We're Not Building Frontier Models. We're Applying Technology to Real Problems.
Charlie Apigian set the framing for Nashville's AI identity before the first session started: this is a community of people who get stuff done. Not frontier model builders, people applying technology to real-world problems through people, process, and domain knowledge.
"Where there is no room for error, trust will arise."
Salisbury's caution about agents rang the clearest: they have confidence, not competence. The $40-50M savings Nexigen generated came from replacing a team manually reviewing data at speed with a system that could not move quickly past errors. The constraint was the point.
Anuj Kapedia at Oak Ridge National Laboratory showed what domain knowledge looks like at its furthest edge: digital twins of humans, built from CT scans, used to test cancer treatments before touching a patient. Moving from reactive to proactive medicine, and bypassing the ethical wall of animal testing.
The Titans AI panel hit a familiar note: fan experience, smart building management, employee productivity. AI applied to a specific place with specific constraints, the Disney parks model, optimizing for presence rather than scale. Brian Moyer's closing line hung in the air: asked how far a solopreneur could take an AI-built product, the answer was "probably farther than you think."
References
Skills Matt Pocock / prototype skill, no URL captured
Community Hack Tennessee / Hackathons International, no URL captured
Community Oracle Academy, free education K-12 and beyond, no URL captured
Design Notes
YIL 2026-05-14 / The Discipline Was Designed by a Human
BMW Corporate provides the structural language: white canvas, dark navy statement bands, rectangular edges, precise section labels, spec-cell metrics, and no drop shadows. The page adds a generated technical hero image to make the human-designed discipline theme visible without leaving the BMW Corporate register.
accent-1 / skills accent-2 / structure accent-3 / decision-science accent-4 / domain-knowledge