Skip to content

SKILL.md

Latest commit

 

History

History
191 lines (132 loc) · 7.67 KB

SKILL.md

File metadata and controls

191 lines (132 loc) · 7.67 KB
name transcript-processor
description Transforms raw interview or podcast transcripts into structured Q/A digests, with optional commentary layer. Use this skill whenever the user provides a transcript (interview, podcast, earnings call, press conference) and wants it cleaned, organized, or analyzed. Trigger words include: "organize this transcript", "clean up this interview", "Q/A format", "digest", "summarize this interview", "what did X say about Y". Also trigger when the user uploads a long conversation text and wants to extract key points or prepare it for sharing. Handles any output language. Always use this skill for transcripts longer than ~15 minutes of content.
tags
claude-skills
transcript-processor
podcast-tools
interview-analysis
when_to_use When the user provides a raw transcript file (interview, podcast, meeting notes) and wants to clean up filler words, structure it into a Q&A format, or generate a deep digest/analysis without losing original context.

Transcript Processor Skill

Modes

This skill has two modes. Infer from the user's request; ask only if genuinely ambiguous.

Mode When to use Output
clean "clean this up", "Q/A format", "文稿清洗", "整理成对话格式" Q + A only, no commentary
digest "digest", "analyze", "what did X avoid", "实话实说", "深度整理" Q + A + Commentary + Meta-analysis

Default to clean unless the user explicitly asks for analysis or commentary.

Before doing anything: state the plan and wait for confirmation

Before running any script or writing any output, state the execution plan in this exact format and stop:

**Plan**
- Mode: clean / digest
- Output language: [e.g. English / Chinese / bilingual: EN transcript, ZH commentary]
- Translation: yes / no — [Q and A will / will not be translated to target language]
- Steps: 1. save transcript → 2. run parser → 3. clean Q/A [→ 4. translate if applicable] → 5. write file

Confirm to proceed.

Do not start executing until the user confirms. This catches mismatches in language, translation intent, and mode before any work is done.

Mandatory second step: run the parser

Before writing any Q/A blocks, always run the transcript parser script.

This is not optional. The script does mechanical timestamp-based segmentation that prevents omissions. Do not rely on reading and remembering — traverse, don't retrieve.

python3 scripts/parse_transcript.py <transcript_file>

The script outputs a numbered skeleton of all segments with timestamps and types (Q / A / QA-mixed). Review the skeleton, correct any misclassifications manually if needed, then proceed to cleaning.

If the transcript file is not yet saved to disk, write it first:

cat << 'TRANSCRIPT' > /tmp/transcript.txt
[paste transcript content]
TRANSCRIPT
python3 scripts/parse_transcript.py /tmp/transcript.txt

Step 1: Clean Q (questions)

Q is cleaned original speech. Remove only:

  • Filler words: uh, um, you know, I mean, like, right (standalone)

Do NOT:

  • Rephrase, reorder, or summarize
  • Merge separate questions into one
  • Remove the interviewer's framing, edge, or follow-up logic
  • Drop the timestamp

Format:

**[timestamp]**

**Q:** Cleaned question text here.

Step 2: Clean A (answers)

A is cleaned original speech — not a summary, not a paraphrase. The speaker's own words must be preserved.

Remove only:

  • Filler words: uh, um, you know, I mean, like, right (standalone)
  • Repeated false starts: "I think I think I think" → "I think"
  • Redundant hedges that repeat within the same sentence

Do NOT:

  • Rephrase, reorder, or substitute words
  • Compress or summarize
  • Drop specific numbers, named products, or named people
  • Editorialize (save that for Commentary in digest mode)

If the answer spans multiple transcript segments, combine them in order.

Step 3 (digest mode only): Write Commentary

Skip entirely in clean mode.

Commentary is independent analysis — not a restatement of A. For each exchange, ask:

  1. Did they answer the question? If not, what did they pivot to, and why?
  2. What structural tension did they avoid? (e.g. answering "is revenue up?" with "costs are down")
  3. What's verifiable? Label claims: [verifiable] / [inference] / [industry consensus]
  4. Does their analogy actually hold? Name it if it doesn't.

Depth: 1-2 sentences for minor topics, full paragraph for core topics.

Step 4 (digest mode only): Meta-analysis

Skip entirely in clean mode.

After all Q/A/Commentary blocks, add a final section identifying recurring patterns:

  • Deflection patterns (e.g. "always cites historical data when asked about future risk")
  • Framing choices (e.g. "consistently reframes competitor gaps as 'different markets'")
  • The 2-3 moments of genuine candor (high-signal, stand out against the baseline)
  • Claims that can be cross-checked against public records

Output language, translation, and section labels

Translation rule

If the user specifies an output language different from the transcript language, translate Q and A into the target language after cleaning. Translation comes after cleaning — clean first in the original language, then translate.

  • "中文 clean" → clean English transcript, then translate Q and A to Chinese
  • "bilingual" → keep original language for Q and A, target language for Commentary only
  • "English clean" on an English transcript → no translation needed

When in doubt, include Translation: yes/no explicitly in the plan step and confirm with the user.

Section labels

Match labels to output language. Do not hardcode any single language.

Section English Chinese Japanese
Question **Q:** **Q：** **Q：**
Answer **A:** **A：** **A：**
Commentary **Commentary:** **实话：** **解説：**
Meta section ## Meta-analysis ## 元分析 ## メタ分析
Verification [verifiable] [inference] [industry consensus] 【可验证】 【推断】 【行业共识】 translate

For bilingual output (e.g. EN transcript + ZH commentary):

  • Q and A: original language label + original language text
  • Commentary: target language label + target language text

Document header:

> Source: [name / date]
> Mode: clean / digest
> Output language: [e.g. English / Chinese / bilingual: EN transcript, ZH commentary]
> Translation: yes / no
> Note: Q and A = cleaned original speech (fillers and false starts removed only)

File output

  • Filename: [interviewee]_[source]_[mode].md — e.g. sundar_allin_digest.md, sundar_allin_clean_zh.md
  • One file per mode/language combination
  • Run scripts/check_digest.py on the output before presenting to the user

Common failure modes

Failure Fix
Skip parser script, go straight to writing Never. Always run parse_transcript.py first
A becomes a summary A is cleaning only — preserve all words except fillers and false starts
Q loses the timestamp Every Q block must have its timestamp
Commentary added in clean mode Clean mode has no commentary. None.
Skipping segments that seem minor Traverse by timestamp — importance is irrelevant at this step
"No direct answer" treated as neutral In digest mode: name what was avoided and why
Meta-analysis missing in digest mode Always required in digest mode, never in clean mode
Output language ≠ transcript language but no translation done If output language differs from transcript, translate Q and A after cleaning
Executing before user confirms plan Always state plan and wait for explicit confirmation first