AI agent skill · by @undefined-ui
Second Brain Transcript
Clean a raw transcript before ingestion: punctuate, paragraph, label speakers, fix mistranscribed technical terms, and split long recordings by topic. Use this skill when the user drops an auto-generated transcript, subtitle file, podcast or lecture transcript, meeting recording text, or asks to process a video or audio source. Do NOT use for text that is already prose, for summarising a transcript, or for the ingestion itself.
Archived 2026-10-23 · Source updated 2026-10-09 · License: MIT. Check the original repository for the latest version.
SKILL.md instructions
# Clean a transcript Concept pages built from unpunctuated caption dumps are noticeably worse, because the model spends its attention reconstructing sentence boundaries instead of understanding claims. ## Core rule Clean, never summarise. Nothing is cut. This is a formatting pass on the record, and the record has to stay faithful. ## Workflow 1. **Punctuate and paragraph.** Sentence boundaries first, then paragraphs at topic shifts. 2. **Label speakers** where they change. In an interview this is most of the value. 3. **Fix technical terms.** Auto-captions mangle proper nouns and jargon consistently. Flag anything you could not resolve rather than guessing. 4. **Keep coarse timestamps** every few minutes, so a claim on a wiki page can be verified in ten seconds instead of a rewatch. 5. **Propose split points** if the recording covers separate topics. A three-hour podcast is not one source. 6. **Write frontmatter:** title, channel or speaker, URL, date. ## Output format The cleaned transcript, then: ``` Cleaned: <source> Speakers labelled: <n> Terms corrected: <list> Uncertain: <list of what could not be resolved> Suggested splits: <topics with timestamps, or none> ``` ## Calibration Never drop filler that carries meaning, such as hedging or a speaker correcting themselves. Those are exactly what distinguish a claim from an aside. If the transcript is too garbled to clean reliably, say so and recommend re-pulling it rather than producing a plausible reconstruction.