classify-post-to-section

name	classify-post-to-section
description	Assigns a substacker draft or published post to the best-fitting section (or to unassigned) based on content + section promises in section-map.md. Used by the Editor on every draft review (to load the right voice overlay) and by the Curator in batch mode. Trigger keywords — classify post, section assignment, which section, route post, per-draft section.

Classify Post To Section

Workflow

Per post (draft or published):
- [ ] Step 1: Read post body (not just title)
- [ ] Step 2: Read section-map.md for all current section promises
- [ ] Step 3: Score post fit against each section's promise (specific, testable, voice register)
- [ ] Step 4: If top score clearly above second → assign that section
- [ ] Step 5: If ambiguous between two sections → propose both; writer picks
- [ ] Step 6: If no section scores above threshold → assign `unassigned`
- [ ] Step 7: Return: {section_slug, confidence, rationale}

Scoring dimensions

For each section, compute fit on:

• Promise match: does this post deliver on the section's one-sentence promise?
• Voice register: does the post's register (confessional-operational / technical-mechanistic / epistolary) match the section's voice overlay?
• Topic tag overlap: does the post's internal topics frontmatter intersect with the section's typical topic distribution?

Worked example

Draft: "KV Cache as a library card catalog" — full body on KV cache mechanics, diagram-heavy, cites Vaswani et al. and Dao et al.

Current sections:

• kalshi-log: scoreboard-required, prediction markets / IPL. Promise match: low. Score: 1/5.
• agent-workshop: mechanism + architecture, code-fence-welcome. Promise match: high. Score: 5/5.

Output: {section_slug: agent-workshop, confidence: high, rationale: "mechanism post with explicit paper citations; matches Agent Workshop register and promise"}.

Guardrails

1. Never assign without scoring at least 2 candidate sections.
2. If writer has manually annotated section: X in frontmatter, respect it — this skill only proposes when frontmatter is missing or unassigned.
3. Ambiguous cases return both candidates; do not arbitrarily pick.
4. "unassigned" is a valid output. Don't force fit.
5. Read body, not just title.