LANVAR
← The shelf

Feynman

Escaped the lab

The opposite of the caveman skill: a Claude skill that teaches like Feynman — the problem first, the naive fix failing, the idea earning its name, then the math worked with real numbers. Trained, not just written, with Microsoft SkillOpt.

Claude skillsSkillOptHTML/SVG 2026
✳ How it actually works
"the movie was great" "the movie was NOT great" letter matcher counts shared words "same!" ✓ opposite meanings — letters can't tell "same!" ✓ opposite meanings — letters can't tell [ 0.9, 0.2 ] [ −0.8, 0.3 ] [ 0.9, 0.2 ] [ −0.8, 0.3 ] far apart = different meaning far apart = different meaning TEACH ME ENCODERS MEANING → NUMBERS THE MEANING MAP NOW the problem — compare a million sentences the naive attempt — match letters. it fails. the idea — turn meaning into numbers: an encoder the math — distance between vectors, worked by hand problem → naive fails → the idea → the math, worked the name "encoder" only appears after the failure that demands it
FIG 1 · one lesson — the problem fools the naive fix, the idea earns its name, the math gets worked
THE LESSON PIPELINE "I don't get X" any concept, any field find the forcing problem what was hard before this thing existed? SIX BEATS, IN ORDER 1 · the problem — no jargon 2 · the naive attempt fails 3 · the idea earns its name 4 · the math, worked 5 · where it breaks 6 · recap card one .html no CDN, no network, yours SKILLOPT — THE LOOP THAT TRAINED THE TEACHER, TWICE roll out on 20 concepts encoders → eigenvalues, one page each LLM judge, 7 criteria problem-first weighs 3×, unworked formulas fail text gradients failures rewritten as edits to the skill.md validation gate no val improvement, no merge skill.md updated the prompt is the weights gen 1 taught it "no jargon in the scene" — gen 2 taught it to work the math ↺ best_skill.md feeds the pipeline ↑
FIG 2 · the machinery — a request in, six beats up, one page out — and the loop that trained the teacher, twice
The story

The caveman skill makes Claude grunt to save tokens. Feynman is its exact inverse: spend the tokens, build the understanding. Ask it to teach you encoders and you don’t get a definition — you get a single interactive HTML page that opens with the problem: a machine has to compare a million sentences by meaning, and the obvious fix, matching letters, calls “the movie was great” and “the movie was NOT great” the same sentence. You watch the naive attempt fail before the word “encoder” is ever allowed on screen. The concept has to earn its name by fixing the exact failure you just saw.

Every lesson walks six beats in order: the problem rendered with zero jargon, the naive attempt failing interactively, the idea emerging as the fix, the core math worked — not displayed — with numbers small enough to recompute by hand, the boundary you can trigger and watch break in real time, and a recap card. The depth contract is the anti-caveman clause: “it can be shown that” is banned; it gets shown. Symbols wear the same colors as the scene objects they stand for, so by the time the equation arrives it reads as a caption of the story so far. One file, inline everything, no CDN.

The interesting part is how the skill got good: it was trained, not just written. Microsoft’s SkillOpt treats the skill markdown as trainable state — the lab built it a custom benchmark of twenty concepts across eight fields (encoders included), and an LLM judge scoring seven weighted criteria where problem-first weighs triple and an unworked formula is a failing grade. The loop rolls the skill out, turns judged failures into natural-language gradients, edits the skill document, and a validation gate keeps only edits that improve held-out concepts. The prompt is the weights.

This is generation two of that loop. Generation one lifted the held-out pass rate from one-in-five to three-in-five and learned rules nobody wrote — like “formal units leaking into a slider’s readout chip count as jargon.” Those rules were folded into the new seed before the whole thing was retrained on the problem-first arc, through the same Claude Code CLI that serves the skill afterwards — and generation two pushed the held-out pass rate from two-in-five to four-in-five, learning along the way that the failing scene and the worked math must share the same numbers, and that a real prediction means committed buttons with consequences, not a slider nobody checks. Caveman grunts. Feynman teaches. The training curve says which one you want explaining eigenvalues.

GitHub ↗
Next on the shelf Agentic SaaS Starter →