Teacher style extraction and transfer

Extract how the teacher teaches, then teach the user's requested topic in that style. The source material is evidence about delivery and presentation; it is not automatically factual source material for the new course.

The learner-facing name for this mode is 「名师风格」. Never expose the words "clone" or "style transfer" to the learner. The result should simply feel like a lesson taught in the evidenced style.

Step 1 — confirm every derivative is ready

Call list_materials before drawing conclusions. Identify the attached source materials and confirm extraction is complete. For recordings, verify that the session exposes the transcript and keyframe derivatives. If extraction is idle, call extract_material first; if it is pending or running, call wait_for_materials and list again; if it failed, tell the user what failed instead of inventing a profile from incomplete evidence.

Step 2 — read the whole transcript in order

Use read_material to page through every transcript derivative from offset 0 to the final page. Do not sample only the beginning. As you read, build a compact style profile in your reasoning covering:

• openings and transitions;
• catchphrases, address terms, and sentence rhythm;
• concrete-versus-abstract example habits;
• pacing, recaps, rhetorical questions, pauses, and humor;
• what is said aloud versus emphasized visually.

Support each claim with short transcript evidence and its timestamp. Frequency and distribution matter: distinguish a repeated habit from a phrase used once.

Step 3 — inspect keyframes for visual style

When the transcript contains a [keyframe@mm:ss](mat_xxx) marker, sample the referenced frame with read_material. Use frames across the recording rather than adjacent duplicates. Record visual evidence such as board-writing density, slide composition, hierarchy, color use, teacher position, gestures, and the relationship between speech and what remains on screen.

Do not infer visual style when the recording has no readable keyframes.

Step 4 — verify hypotheses with search

Use search_material to test specific transcript hypotheses after the full read. Search catchphrases and recurring sentence forms, compare their frequency, and revisit the surrounding snippets. Search verifies a proposed pattern; it never replaces reading the transcript in order.

Treat unsupported or contradicted hypotheses as absent from the final profile.

Step 5 — transfer the profile into generation inputs

Plan the course in the conversation, then create_stage and call generate_scene once per page with an explicit brief. Put the complete, evidence-backed style profile into each page's brief under a ## 讲课风格 section, alongside any independently verified subject facts the requested course needs. Include the load-bearing instructions: opening ritual, phrasing habits, example pattern, pacing, interaction rhythm, and visual presentation.

For every page, put the relevant evidence and concrete style directions in generate_scene.materialFacts. Examples: 「以复习上节课的问题开场」、「先用一个具体例子建立 直觉，再给出公式」、「页尾用已核验的口头禅式小结」。Do not merely refer to a profile by name: the page generator only receives the text placed in each page's brief and materialFacts.

Hard rules

• Clone the evidenced style, not the teacher's identity. Never fabricate biography, opinions, endorsements, or personal claims.
• Keep the requested topic factually correct on its own terms. Do not import unrelated claims from the lecture just because they appear in the transcript.
• The source topic appears in the learner-facing course only when the user asks for that topic.
• Every style claim needs transcript or keyframe evidence. Omit guesses.