transcript-cleanup

Transcript Cleanup

What This Skill Does

Cleans and formats a raw auto-generated podcast transcript for publication on a website or show page — removing filler words and transcription errors, adding speaker labels, formatting for readability, and preserving the authentic voice of host and guests.

When To Use This Skill

• You have a raw transcript from an auto-transcription service (Otter.ai, Descript, Whisper, Rev) and need it cleaned for publication
• You want to publish an accessible transcript on your show's website alongside the episode
• Your sponsor or network requires a clean transcript for accessibility compliance
• You want a readable written record of an interview that can be shared as a standalone document

What You Need To Provide

Required:

• The raw transcript text (paste directly)
• The names of the host and guest(s), so the assistant can assign correct speaker labels

Optional:

• Show name and episode title (for the header)
• Any proper nouns, technical terms, or names that the transcription service consistently got wrong (the assistant will look for these and correct them)
• Publication format: website (markdown), plain text, or formatted document
• Whether you want light cleanup (fix errors, add speakers, remove obvious filler) or heavy cleanup (also remove all filler words, tighten sentences, remove false starts)

How the Assistant Approaches This

1. Assigns correct speaker labels throughout, replacing auto-generated labels (SPEAKER 1, SPEAKER 2) with actual names
2. Corrects transcription errors: proper nouns misspelled, technical terms mangled, homophones chosen incorrectly, sentences that clearly run together or are incorrectly split
3. Applies the appropriate level of cleanup:
   • Light: Fix errors and add speakers only; preserve all speech patterns including filler words and false starts — this is the most authentic record
   • Heavy: Also remove filler words (um, uh, you know, like used non-meaningfully), clean false starts (incomplete sentences immediately corrected by the speaker), and tighten repeated false starts into the completed sentence
4. Adds minimal formatting: timestamps every 5–10 minutes, a header with episode title and date, clean paragraph breaks between speakers
5. Does not change the meaning, reorder ideas, or silently alter what was said — any uncertain transcription is marked [UNCLEAR] rather than guessed at
6. Closes with a "Next Step" note: how to handle any [UNCLEAR] passages (check against the audio file), and whether show-notes-generator or episode-summary-writer should be run using this cleaned transcript as the source material