Training Lecture Recap

---
name: training-lecture-recap
description: "Turn a video file or transcript into a reader-friendly lecture recap structured into five fixed sections — Summary, Theme, Q&A, Tools, Call to Action — while strictly avoiding any reference-outline detail leakage."
category: training
project: training-lecture-recap
platform: claude
status: Done
author: Peter Tu
sheetId: "4.08"
originalName: "處理影片回顧"
input: "A Fireflies meeting (id/URL/title — preferred, already transcribed), a video file, a video URL, or the verbatim transcript of a recorded lecture/livestream."
process: "Transcribe the video (if not already a transcript), remove filler words, map the cleaned content to the reference outline structure, then render the recap"
output: "A reader-friendly recap document with five sections: Summary, Theme, Q&A, Tools, Call to Action"
synergy: []
house-style: bound

---

# Lecture Recap Writer

```bash
npx skills add https://github.com/peter-tu-zynkr/zynkr-skill-builder --skill training-lecture-recap
```

Turn a recorded lecture, livestream, or workshop into a publishable five-section recap. Use this skill when you have a video or transcript and need a reader-friendly summary in the canonical Zynkr recap format. **Different from sibling skills**:
- `training-lecture-transcript` polishes the raw transcript itself (preserves every line).
- `training-process-video` produces four parallel workstreams (slides, social posts, etc.) — not a single doc.
- `training-lecture-recap` (this skill) produces **one** narrative recap doc with five fixed sections.

---

## Step 1 — Collect inputs

Ask the user for:
1. **Source** — one of:
   - **A Fireflies meeting (prefer this)** — a meeting id, an
     `app.fireflies.ai/view/<id>` URL, or just the title and roughly when it ran
     ("上週那場 Claude Code 講座"). Anything Fred sat in on is already transcribed,
     so this skips the whole transcription step below. Resolve with
     `mcp__fireflies__fireflies_search(query="keyword:\"Claude Code\" from:2026-09-01")`
     and confirm the row by title + date before fetching.
   - Video file (`.mp4`, `.mov`, `.m4a`, `.wav`)
   - Video URL (YouTube, Vimeo, etc.)
   - Transcript text (paste or file path)
2. **Lecture context** — speaker name, topic, target audience (1–2 sentences)
3. **Recap outline** (optional) — a reference structure or session outline if available

Store as `SOURCE`, `CONTEXT`, `OUTLINE_REF` (may be empty).

---

## Step 2 — Get a clean transcript

Branch by input type:

- **Fireflies meeting** → `mcp__fireflies__fireflies_get_transcript(transcriptId="<id>")`.
  Returns speaker-labelled, timestamped sentences already — no ASR pass needed, so
  go straight to the cleanup below. `mcp__fireflies__fireflies_get_summary` also
  gives keywords and action items, handy for the Q&A and Call-to-Action sections.
  ⚠️ Free plan = **50 API requests per day**, one per call — search once, fetch once.
  ⚠️ `audio_url` / `video_url` come back empty on the free plan; text only. If the
  recap genuinely needs the media, fall back to the Video file branch.
- **Already a transcript** → proceed to Step 3
- **Video file** → transcribe locally or via the user's preferred ASR tool (suggest `whisper` or `training-lecture-transcript` skill for the cleanup pass)
- **Video URL** → ask the user to download or provide the transcript; this skill does not fetch external media

Once you have raw transcript text:
1. Remove obvious filler words (嗯、啊、那個、就是, "um", "uh", "like")
2. Normalize speaker labels if present
3. Preserve direct quotes from the speaker — these become Q&A material

Store as `CLEAN_TRANSCRIPT`.

---

## Step 3 — Map content to the five sections

Walk through `CLEAN_TRANSCRIPT` and bucket content into the five canonical sections. Strict rule: **if a reference outline was provided in `OUTLINE_REF`, use it only to structure section order — never copy detail text from the outline**. Detail comes from the transcript only.

### Section structure

**1. Summary (摘要)** — 2–3 sentences capturing the lecture's central message. Reader-friendly, third-person.

**2. Theme (主題)** — The lecture's argument or framework. 3–5 bullets, each one main point with a one-sentence explanation drawn from the transcript.

**3. Q&A** — Either:
   - Actual audience Q&A from the recording, OR
   - Inferred Q&A: pull 3–5 questions a reader would likely ask, then answer them using direct quotes from the speaker.
   Format: `Q: ... / A: <quote or paraphrase from the lecture>`

**4. Tools (工具)** — Concrete tools, frameworks, books, or references the speaker mentioned. Bullet list with one-line context per item.

**5. Call to Action (行動呼籲)** — 1–2 sentences. What does the lecture invite the reader to do next? Pull from the speaker's actual closing remarks if present; otherwise infer the most natural reader action.

---

## Step 4 — Output

Render in zh-TW unless the lecture is in English (then English). Use this template:

```
# 講座回顧：<lecture title>

> Speaker: <name>  ·  Audience: <audience>

## 摘要 / Summary
<2–3 sentences>

## 主題 / Theme
- **<point 1>** — <explanation>
- **<point 2>** — <explanation>
- **<point 3>** — <explanation>

## Q&A
**Q:** <question>
**A:** <answer drawn from transcript>

(repeat 3–5 times)

## 工具 / Tools
- **<tool name>** — <one-line context>

## 行動呼籲 / Call to Action
<1–2 sentences>
```

---

## Step 5 — Quality checks before delivery

Before handing the recap to the user, verify:
- [ ] Summary is under 80 words
- [ ] Theme has 3–5 bullets, no overlap with Summary
- [ ] Q&A pulls from actual transcript content, not invented
- [ ] No detail copied verbatim from `OUTLINE_REF`
- [ ] All speaker quotes are accurate (paraphrase if you can't preserve verbatim)
- [ ] CTA is concrete and actionable, not generic

---

## Rules

- Never invent statements the speaker did not make
- Never copy reference outline text into the body — outline is for ordering only
- Always preserve the speaker's voice — if they're casual, the recap is casual; if formal, formal
- Filler-word removal is editorial — do not change meaning
- When uncertain about a fact mentioned in the lecture, mark with `[unclear in source]` rather than guessing

## House style

Writing style is **not owned by this file**. The house voice lives in two Google Docs under
`[@] 寫作指南` (`12DBdFz3SK22ie9im_ThFMI7IBRXsTZsV`), read at runtime:

- 《[2.0] Zynkr 通用風格指南 House Voice》 `10bOIQwRm9Pxwgct4hlwCwK_B4Pipai1HqBPZKzyRHSE` —
  the universal core, plus the addendum for this surface
- 《[3.2] 禁用詞清單 Forbidden Words》 `1N5sHLP4qzmmhpCGsi6KElxi1z0MFe4QZ0Q_35T10Uyg`

Read both before producing client- or reader-facing text, and scan the draft against 《[3.2]》
before handing it over. If Drive is unreachable, say so in the output rather than proceeding
unchecked. Never re-implement either list inside this file.