Insights

You don't need forty education skills. A working AI tutor is four composable behaviors — question, explain-and-ground, drill, and assess — assembled from the marketplace. Here's the stack.

LLMs are extraordinary analogy engines — the core of a good explainer. They're also confidently wrong in ways a novice can't catch. Great AI explanation is analogy plus grounding, never analogy alone.

An AI that rewrites the student's essay teaches nothing. The skilled move is rubric-based feedback that diagnoses without doing the work — preserving the struggle where the learning actually lives.

A hallucinating tutor is worse than no tutor — it teaches errors with authority. Grounding a tutor in real tools (a math engine, references, a code runner) is a safety requirement, not a nice-to-have.

The durable win from an AI tutor isn't the physics answer — it's teaching the learner how to plan, self-test, and manage anxiety. Metacognition, not content, is what compounds.

Most ed-tech fads died. Spaced repetition — testing yourself at widening intervals against the forgetting curve — didn't. Here's why it's the highest-leverage thing to wire an AI tutor into, and the skills that do it.

Bloom's 2-sigma problem — that one-on-one tutoring beats classroom instruction by about two standard deviations — was a scaling impossibility for 40 years. Personalized AI tutors are the first plausible answer, with real caveats.

How a solo builder plus an AI agent ship on-brand UIs fast using a DESIGN.md from the registry. The leverage that closes the design gap without a designer.

Every model launch ships with bar charts where the new model wins. Here's a reusable checklist for sanity-checking those numbers — using Grok 4.5's July 2026 launch as the worked example.

Grok 4.5 ranks #4 on the independent Intelligence Index at a price >60% below Opus 4.8 and GPT-5.5. Here's what that price-per-intelligence unlocks for high-volume skill and agent workloads.

xAI reports Grok 4.5 using ~4x fewer output tokens than Opus 4.8 on SWE-Bench Pro. For long agentic loops where output dominates the bill, that can matter more than winning the benchmark.

'On par with GPT-5.5/Codex at lower cost' sounds decisive — until you notice it's a harness score, not a model score. What that means when you pick a coding backend for a skill.