second-brain-transcript

1# Clean a transcript
2 
3Concept pages built from unpunctuated caption dumps are noticeably worse,
4because the model spends its attention reconstructing sentence boundaries
5instead of understanding claims.
6 
7## Core rule
8 
9Clean, never summarise. Nothing is cut. This is a formatting pass on the record,
10and the record has to stay faithful.
11 
12## Workflow
13 
141. **Punctuate and paragraph.** Sentence boundaries first, then paragraphs at
15   topic shifts.
162. **Label speakers** where they change. In an interview this is most of the
17   value.
183. **Fix technical terms.** Auto-captions mangle proper nouns and jargon
19   consistently. Flag anything you could not resolve rather than guessing.
204. **Keep coarse timestamps** every few minutes, so a claim on a wiki page can
21   be verified in ten seconds instead of a rewatch.
225. **Propose split points** if the recording covers separate topics. A
23   three-hour podcast is not one source.
246. **Write frontmatter:** title, channel or speaker, URL, date.
25 
26## Output format
27 
28The cleaned transcript, then:
29 
30```
31Cleaned: <source>
32Speakers labelled: <n>
33Terms corrected: <list>
34Uncertain: <list of what could not be resolved>
35Suggested splits: <topics with timestamps, or none>
36```
37 
38## Calibration
39 
40Never drop filler that carries meaning, such as hedging or a speaker correcting
41themselves. Those are exactly what distinguish a claim from an aside.
42 
43If the transcript is too garbled to clean reliably, say so and recommend
44…