Topic Splitting

Long sources should usually be split before they are built into artifacts.

Meetings, recordings, interviews, and long articles often contain mixed signal: some local context, some noise, and a few reusable rules. split-source divides the source into topic candidates and scores each candidate with the same readiness gate used by analyze.

Command

source-to-skill split-source out/recording-clean.md --out out/topics

This writes:

• topic-report.md with score and recommended output per segment
• segments/01-*.md, segments/02-*.md, and so on

For integrations:

source-to-skill split-source out/recording-clean.md --out out/topics --json

Build Or Fold A Segment

After reviewing topic-report.md, build only the segment that has enough signal:

source-to-skill build-segment out/topics 2 --level seed --out out

Or fold it into an existing skill:

source-to-skill fold-segment out/topics 2 ~/.codex/skills/design-review

The index is the 1-based segment number shown in topic-report.md and encoded in the segment filename, for example segments/02-useful-rule.md.

When To Use

Use splitting before building when the source is:

• a meeting transcript
• a voice note
• an interview
• a long article with multiple sections
• a document that mixes project context and reusable method

Review Rule

Treat each segment independently. A strong source can contain weak segments, and a weak meeting can still contain one useful seed.

Do not build from the whole source just because one segment is valuable. Build from the segment whose own score, evidence, and scope support the artifact.