Generate transcript → Generate review draft (topic-level outline + sentence-level deletion suggestions) → User review → Execute editing

• Audio/video files
• (Optional) List of speaker names, e.g.,

  ["Maia", "响歌歌", "安安"]

1. Review Draft - A file containing: content outline + full transcript + deletion marks
2. After Confirmation - Execute editing

• Total sentences: 3390
• Suggested deletions: 377 instances
• Silence segments: 3 instances (01:22)

• Opening small talk: 31 instances
• Chit-chat - personal background: 23 instances
• Technical debugging: 15 instances
• Recording discussion: 6 instances
• Privacy - company name: 5 instances
• Privacy - location: 4 instances
• Privacy - school name: 3 instances

/podcastcut-content

/podcastcut-transcribe

• Sentence-level transcription + speaker diarization = accurate speaker identification
• Character-level transcription + speaker diarization = easy speaker misalignment (OOM for long audio, alignment difficulties after segmentation)
• First delete large segments (sentence-level), then process the remaining content (character-level)

• Semantic off-topic/chit-chat (no obvious keywords)
• Chit-chat after guest introduction (where they live, graduation year, school conditions)
• Hidden recording discussions (no keywords like "cut")

1. Opening small talk: Chit-chat, technical debugging before the official opening ("Hello everyone")
2. Recording discussion: Discussion of editing, recording status, technical issues, "Should we cut this segment"
3. Privacy - company name: Mention of specific company names (Google, Meta, ByteDance, etc.)
4. Privacy - school name: Mention of specific school names (Stanford, Tsinghua, etc.)
5. Privacy - location: Mention of specific locations (Palo Alto, Silicon Valley, etc.)
6. Privacy - personal name: Mention of specific personal names (non-public figures)
7. Off-topic/chit-chat: Discussion irrelevant to the podcast topic (personal background chit-chat, geographic discussion, etc.)
8. Redundant repetition: Repeating the same point multiple times, large segments of repetition

• Sentence timestamp
• Deletion type
• Reason (brief description)

• Sudden appearance of place names, school names, years
• "Which area do you live in" "When did you arrive" "How is it over there"
• Multiple consecutive sentences discussing non-topic content (geography, school, city comparison)

• Technical issues: "Can you hear me", "Disconnected", "Headphone battery dead"
• Content concerns: "Too low-key", "Don't want to share", "Don't mention details"
• Editing discussion: "Cut later", "Should we keep this"

1. The review draft is the final source for user review
2. Before executing editing, re-parse deletion marks from the review draft
3. Do not rely on the potentially outdated

   podcast_删除清单.json

~~strikethrough~~

/podcastcut-content     → Content editing (semantic level) ← This Skill
/podcastcut-edit        → Execute editing
/podcastcut-transcribe  → Verbal error identification (technical level, optional)
/podcastcut-subtitle    → Generate subtitles

Original video
    ↓
/podcastcut-content  ← Mark large segments (small talk, off-topic, redundancy, privacy)
    ↓
/podcastcut-edit     ← Execute deletion, output v2
    ↓
【Optional】Need to process verbal errors?
    ↓ Yes
/podcastcut-transcribe  ← Identify verbal errors, filler words, silence
    ↓
/podcastcut-edit        ← Execute deletion, output v3
    ↓
Completed

• After deleting large segments, the video becomes shorter
• Verbal error identification transcription is faster, review scope is smaller
• No need to process verbal errors in deleted large segments

• Speaker ID misalignment caused by segmented transcription: Speaker IDs reset independently for each segment during segmented transcription, resulting in different IDs for the same person after merging
  • Cause: 2-hour audio split into 13 segments to avoid OOM, each segment's speaker ID starts from 0
  • Solution: Prioritize full transcription (2-hour audio takes about 16 minutes, no OOM)
  • Updated: Added "Segmented vs Full Transcription" section to tips/转录最佳实践.md
• FunASR may identify the same person as multiple IDs: 3-person conversation identified as 4 speaker IDs
  • Performance: 响歌歌 was split into spk1 (60 sentences) and spk3 (560 sentences)
  • Solution: Search self-introduction phrases ("I'm host xxx") to confirm mapping, merge multiple IDs
  • Updated: Added confirmation method and common issues to the "Speaker Diarization" section of SKILL.md
• Updated performance data: 2-hour podcast tested to take 16 minutes, 3390 sentences (previous estimate of 12 minutes, 800 sentences was low)
  • Updated: Performance reference table in tips/转录最佳实践.md

• Incomplete review draft content: AI took a shortcut and wrote "(The following content is topic discussion, keep...)" instead of the full transcript
  • Updated: Clearly marked in the "Section IV" of the review draft format that full content must be output, no omissions allowed
• Redundant transcript file output: Users only need one review draft (which already includes the full transcript)
  • Updated: Removed

    podcast_逐字稿.md

    from the output files section, clearly stated to only output the review draft

• Deleted "User Confirmation Method" section: Users actually operate by directly modifying deletion marks in the review draft, no need for command-based operations
  • Old workflow: Users input commands like "Delete topics 1, 3" "Delete all silence"
  • Actual workflow: Users add/remove

    ~~strikethrough~~

    in the review draft, then say "Cut according to the review draft"
  • Updated: Flowchart, example dialogue, removed command-based operation instructions
• FunASR call parameter error caused transcription failure: Using simplified model names cannot obtain

  sentence_info

  • Incorrect writing:

    model="paraformer-zh"

    +

    spk_model="cam++"

    +

    sentence_timestamp=True

  • Correct writing: Must use full model path + VAD + Punc + Speaker four models
  • Cause: SKILL.md only wrote simplified parameters, actual execution used incorrect API due to "free play"
  • Updated: Included complete call code in SKILL.md, marked comparison between incorrect and correct writing
  • Lesson: Executable code must be fully written in SKILL.md, cannot only write parameter names and let AI assemble it

• Reverted to sentence-level timestamps: Character-level + speaker diarization is unstable for long audio
  • Issue: After character-level transcription and merging speaker information, speaker alignment errors occurred ("I'm host 麦雅" was attributed to 响歌歌)
  • Cause:
    1. Speaker diarization OOM for long audio (2-hour audio → 234MB WAV)
    2. Segmented speaker diarization returned 0 sentences (API format issue)
    3. Character-level transcription has no punctuation, sentence boundaries are unnatural
  • Updated: Reverted to sentence-level timestamps, this Skill only deletes entire sentences
  • Half-sentence deletion is left to

    /podcastcut-transcribe

    (character-level)

• Attempted to upgrade to character-level timestamps: Solved the problem that sentence-level cannot accurately delete parts of sentences
  • Issue: Deleting "Um, I can talk about why" would also delete the second half "This episode's guest was actually invited by 响歌歌"
  • Attempt: Use 30s segmentation +

    timestamp_granularity="character"

    to obtain character-level timestamps
  • Result: Character-level transcription succeeded, but speaker diarization failed, leading to speaker alignment errors
  • Final decision: Revert to sentence-level (see 2026-01-25 feedback)

• Added silence detection function: Use FFmpeg silencedetect to identify long blank segments (≥3 seconds), listed separately in the review draft for user confirmation to delete
• Review draft marked for deletion but not actually cut: Entire sentences marked for deletion in the review draft, but only part of the content was in the deletion list
  • Case: Review draft

    ~~Um, this is why specifically...~~

    , deletion list only had

    Um,

  • Updated: Emphasized that the review draft and deletion list must be synchronized, re-parse from the review draft before executing editing

• Adjusted transcript/review draft format: Content from the same speaker is concatenated, no line breaks per sentence
  • Original format: One line per sentence
  • New format: All sentences from the same speaker are in one paragraph
  • Advantage: More compact, better reading experience

• Must use Claude for semantic analysis: Rule-based keyword matching is not sufficient in quality
  • Issue: Cannot identify semantic off-topic/chit-chat, hidden recording discussions
  • Updated: Added "AI Analysis Method" section, clearly stated that Claude must be used to analyze the transcript in segments
  • Included: Analysis workflow, Claude prompt template, detailed deletion type explanations

• Adjusted format of Section II of the review draft: Changed to full transcript + inline deletion marks
  • Original format: Grouped by topic → deletion suggestions listed under each topic (table form)
  • New format: Full transcript (main text), deleted content marked with

    ~~strikethrough~~

    +

    [Delete: Reason]

    inline
  • Advantage: Retains full context, users can see the context before deciding whether to delete

• Large segment deletion not clean: Consecutive sentences marked for deletion, but sentence-by-sentence deletion retained blank spaces between sentences
  • Cause: Each sentence processed independently, no merging of consecutive deletions with the same reason
  • Updated: Clarified division of labor, this Skill focuses on sentence-level, filler words/verbal fillers processed by

    /podcastcut-transcribe

  • Editing rules have been synchronized to

    /podcastcut-edit

• Do not mark filler words at sentence-level: Sentence-level timestamps are not precise enough, deleting filler words is easy to cause accidental deletion
  • Updated: Divided deletion types into "sentence-level" and "character-level", clarified division of labor

• Recording discussions and technical debugging may appear anywhere in the podcast, not just at the opening
  • Updated: Recording-related detection changed to full-process detection, added technical issue keywords and continuous paragraph detection
• Chit-chat after guest introduction (where they live, when they arrived, how the school is) is easy to miss
  • Updated: Added signal patterns (place names, school names, year-related conversations) to off-topic/chit-chat detection
• Sentence-by-sentence review is inefficient, users want to see the global structure and delete entire blocks
  • Added: Review draft integrates topic-level outline and sentence-level deletion suggestions, complete review in one file