Files
reliquary/plugins/spire/skills/curator/SKILL.md
T
2026-09-17 14:00:22 +02:00

2.0 KiB

name, description, disable-model-invocation
name description disable-model-invocation
curator Compress accreted refinement sections in workpad documents. Takes a file, reads its dated refinement layers, and consolidates them into a single compressed section. Human-triggered only. true

Curator

Accretion is how the lab remembers; curation is how it stays readable. Igor appends dated refinement sections. This skill compresses them when the layers grow unwieldy — distilling without discarding.

Method: compress-accretions

Input: a file path containing one or more dated refinement sections.

Procedure:

  1. Read the target file. Identify all sections that match the accretion pattern — dated headings added by Igor (typically ### Refined <YYYY-MM-DD> or similar dated subsections appended below original content).
  2. For each cluster of accreted sections belonging to the same parent section:
    • Extract the substantive corrections and additions.
    • Discard superseded content (where a later refinement contradicts an earlier one, keep only the later).
    • Preserve attribution: note the date range covered.
  3. Replace the individual dated sections with a single section:
    ### Compressed <YYYY-MM-DD>
    <!-- Covers refinements from <earliest-date> to <latest-date> -->
    
    The compressed body carries the net effect of all layers — same information, fewer words.
  4. Do not touch the original (non-accreted) text. The author's voice stays intact.
  5. Commit the result. The pre-compression state is recoverable via git diff or git log -p.

Safety:

  • This method rewrites append-only sections. The only rollback mechanism is git history. Run it deliberately, not automatically.
  • When in doubt about whether a refinement is superseded, keep it. Losing a correction costs more than an extra paragraph.

Scope

This skill operates on individual files — workpad documents, adept CLAUDEs, skill files. It does not touch the wiki or knowledge layer; that is Tome's domain.