transcribe-refiner（转录精炼机）

Transcribe Refiner - Caption Cleanup Engine

Transform raw auto-generated captions into clean, readable transcripts with zero content loss.

Core Purpose

Auto-generated captions (Zoom, YouTube, Teams, etc.) are messy: fragmented sentences, timestamps everywhere, speaker tags on every line, filler words, transcription errors. This skill reconstructs them into coherent, flowing text that can be consumed by humans or downstream skills (like lecture-alchemist).

Critical Rules

Zero Content Loss

Every substantive statement, technical term, concept, question, and answer from the raw captions MUST appear in the output. Only noise is removed, never content.

Remove: Timestamps, redundant speaker tags, filler words (um, uh, basically, right?, you know), technical interruptions ("can you hear me?", "let me share my screen"), duplicate sentences from reconnection.

Preserve: Every teaching point, code reference, question asked, answer given, tangent with value, name, URL, command, or technical term.

Smart Error Correction

Auto-captions make predictable errors. Fix them using domain context:

When uncertain about a correction, keep the original and flag it: [unclear: "original text"]

Speaker Handling

• Identify unique speakers from tags
• Normalize names (e.g., [rishabh] → **Rishabh:**)
• Only include speaker attribution at natural conversation changes
• For single-speaker lectures, omit speaker tags entirely after initial identification
• For Q&A, clearly mark: **Student:** and **Instructor:**

Input Formats

Processing Steps

1. Strip noise - Remove timestamps, sequence numbers, formatting artifacts
2. Merge fragments - Join broken sentences across caption blocks
3. Remove filler - Strip "um", "uh", "basically", "right?", "you know" (but keep if they carry meaning like "right?" as a genuine question)
4. Fix transcription errors - Use domain context to correct obvious misrecognitions
5. Remove technical interruptions - "Can you hear me?", "Let me share my screen", "Is my screen visible?", connection issues
6. Form paragraphs - Group related sentences into natural paragraphs by topic
7. Identify sections - Insert --- breaks at major topic transitions
8. Normalize Q&A - Clearly separate questions from instruction
9. Add metadata header - Speaker(s), estimated duration, domain detected

Output Format

# Transcript: [Topic/Title if identifiable]

**Speaker(s):** [Name(s)]
**Estimated Duration:** [from timestamp range]
**Domain:** [Auto-detected: WebDev / AI-ML / Web3 / DSA / General]
**Cleaning Notes:** [e.g., "Fixed 12 transcription errors, removed ~45 filler instances"]

---

[Clean, flowing paragraphs organized by topic]

[Natural paragraph breaks at topic changes]

---

[Next topic section]

---

## Q&A Segments

**Student:** [Question]

**Instructor:** [Answer]

Topic Inventory (Anti-Loss System)

This is the critical mechanism that prevents data loss across the pipeline. After cleaning, generate a Topic Inventory at the end of output -- a manifest of every substantive item found in the transcript.

## Topic Inventory

### Concepts Mentioned
1. [Concept] - paragraph [N]
2. [Concept] - paragraph [N]
...

### Technical Terms Introduced
- [term]: first mentioned in paragraph [N]
...

### Code/Commands Referenced
- [code snippet or command] - paragraph [N]
...

### Questions Asked (Q&A)
- Q: [question summary] - paragraph [N]
...

### Names/Resources Mentioned
- [name, URL, tool, book, etc.]
...

### Corrections Applied
| Original Caption | Corrected To | Confidence |
|-----------------|-------------|------------|
| "lowest function" | "loss function" | High |
| "epic" | "epoch" | High |
| [unclear text] | [kept as-is] | Low |

### Stats
- Raw caption blocks: [N]
- Substantive paragraphs produced: [N]
- Filler instances removed: [N]
- Transcription errors corrected: [N]
- Uncertain corrections flagged: [N]

This inventory travels to the next stage (lecture-alchemist) for cross-verification. Every item in this inventory MUST appear in the final notes.

Timestamp Anchors

Preserve approximate timestamps as hidden anchors for key topic transitions. Format:

<!-- T:20:36:30 --> Neural network architecture introduction
<!-- T:20:45:12 --> Activation functions
<!-- T:21:03:45 --> Training loop

These allow the reader to jump back to the recording at specific points.

Quality Checklist

Before output, verify:

• Every teaching point from raw input is in the output
• Topic Inventory is complete and accurate
• Transcription errors corrected using domain context
• Uncertain corrections flagged with [unclear:...]
• Filler words removed without losing meaning
• Sentences properly merged (no mid-word breaks)
• Q&A segments clearly separated
• Technical interruptions removed
• Timestamp anchors placed at topic transitions
• Output reads as natural, flowing text

Pipeline Position

This skill is Stage 1 in the lecture processing pipeline:

1. transcribe-refiner (this) → clean transcript + Topic Inventory
2. lecture-alchemist → structured study notes (verifies against inventory)
3. concept-cartographer → visual diagrams (verifies against inventory)
4. obsidian-markdown → Obsidian vault formatting

Downstream Packaging Contract

1. Keep source-rich traceability in pipeline artifacts (segment_ledger, coverage_matrix, enhanced_notes).
2. Learner-facing final tutorial note should be sanitized for readability (no inline [source:...] tags).
3. Final published naming should follow:

- <Domain> Class <NN> [DD/MM/YYYY] - <Topic> (title/H1) - <DomainFile> Class <NN> [DD-MM-YYYY] - <Topic>.md (published filename)