AgentSkillsCN

Refine

精炼

SKILL.md

Refine Loop Skill

Autonomous research loop for enhancing openloop notes during the 00:00-04:00 window.

Triggers

  • refine — Start the autonomous refinement loop (runs for one Minimax cycle / 4 hours)
  • refine --note <slug> — Refine a specific openloop note (default timeout: 20 min)
  • refine status — Check current state without running
  • refine stop — Gracefully halt any running loop

Timing Window

Allowed window: 00:00 - 04:00 (one Minimax cycle)

  • Outside this window: refine returns "Too early/late, come back between 00:00-04:00"
  • Inside this window: Loop runs for 4 hours (until 04:00 max)
  • Single note refinement (--note): Always allowed, no time gate

Single Note Refinement

When targeting a specific note (refine --note <slug>):

  1. Load Note — Fetch the specified openloop note
  2. Analyze — Identify open questions and research needs
  3. Research — Answer the primary question (one at a time)
  4. Update — Add findings to the note
  5. Log — Record action in refinement-log.md
  6. Complete — Exit after one successful research cycle

Default timeout: 20 minutes (user can override)

Loop Steps

  1. Select File — Pick an openloop note needing refinement (skip if last_refined is today)
  2. Analyze — Identify open questions not yet researched
  3. Research — One question at a time (API search → search skill → browser fallback)
  4. Update — Add findings to the note
  5. Mark Complete — Set last_refined: <ISO date> in the note's frontmatter
  6. Log — Record action in refinement-log.md
  7. Loop — Return to step 1 or stop

Rules

  • No duplicate refinement: Skip notes with last_refined set to today's date
  • Track refinement: Every refined note gets last_refined: <ISO date> in frontmatter

Morning Summary

At the end of the loop (around 04:00), send JB a Telegram message with:

  • Which notes were refined (list of note titles/slugs)
  • What new knowledge appeared (summary of research findings added to each note)

Anchor: [[skill:refine]]