A skill distilled from 478 ML lectures and the professor who gave them

What it does

hung-yi-lee-skill is a system prompt—or “skill”—that teaches an LLM to explain machine-learning concepts in the exact style of Taiwanese professor Hung-yi Lee. It is not a vague persona wrapper; it is a structured knowledge base built from 478 YouTube lectures, 27 full transcripts, a 916-node concept graph, and a real interview with the professor himself.

The interesting bit

The author treated prompt engineering like fieldwork. They used Fable 5 to statistically mine verbal tics—discovering that Lee says “比如說” 609 times, fourteen times more often than “舉例來說”—then generated an interview protocol to fill in the “negative space” transcripts could not capture. When the real professor contradicted the inferred rules, his answers overrode the data. It is essentially a persona prompt that passed a reality check with the human it impersonates.

Key highlights

• Built from 478 video transcripts and distilled into a 916-node knowledge graph with 3,664 edges and 10 concept communities.
• Enforces a specific teaching architecture: intuition first, then the black box, then the mechanism, then pitfalls, then a recap.
• Verbal DNA is quantified, not guessed; signature phrases and hedges were extracted via frequency analysis rather than intuition.
• The professor himself reviewed the output, and his first-person corrections are baked into SKILL.md as the highest-authority layer.
• Ships with CLI tools and an interactive HTML graph for navigating the underlying concept map.

Verdict

Grab this if you are building non-trivial LLM personas and need a blueprint for going beyond shallow mimicry, or if you want a Mandarin-speaking tutor for deep learning. Skip it if you do not read Chinese or have no interest in Professor Lee’s specific curriculum.