improve-system

The self-improvement loop — parse freshly synced /raw data, auto-apply low-risk fixes, write high-stakes proposals to output/review_[date].md for sign-off. Use for the scheduled routine or "improve th

SKILL.md

GitHub

Improve System

Parse freshly synced raw data and turn it into workspace improvements, segregated into three explicit risk tiers. This is the command the System Improvements routine fires (Tuesday and Friday afternoons, after /data-ingestion).

Step 1 — Process pending reviews first

Before proposing anything new, look for previous output/review_*.md files containing checked boxes:

  • [x] Approve → apply that change now, then mark the item (applied YYYY-MM-DD)
  • [x] Reject → mark it (dismissed) and never re-propose it
  • Answered context questions → fold the answer into this run's analysis

Step 2 — Analyze fresh data

Read output/.improvement-log.md for the last run date (if missing, this is the first run). Scan wiki/index.md for entries added since then, opening raw files only where the index summary suggests a lesson. Look for:

  • Recurring mistakes across session digests
  • Manual workflows repeated 2+ times (skill candidates)
  • Wiki entries contradicted by newer curated content
  • Data bloat: unindexed raw files, duplicate entries, dead links
  • Skills/config that session evidence shows are misfiring

Step 3 — Segregate into tiers

Tier 1 — Auto-approve (apply quietly now): Low-risk hygiene only — indexing unindexed raw files, fixing dead wiki links, deduplicating index rows, linking matching files, tidying data bloat. Never anything under .claude/ or CLAUDE.md. Append each applied change as one line to the hidden log output/.improvement-log.md (YYYY-MM-DD | auto | <what> | <why>).

Tier 2 — Needs sign-off (propose, don't touch): Changing CLAUDE.md/system prompts, modifying or creating skills, altering routines or settings, deleting anything. Write each to today's review file output/review_YYYY-MM-DD.md:

### Proposal N: <title>
- **What**: <exact change, with file paths>
- **Why**: <evidence from raw/wiki, with links>
- **Risk**: <what could break>
- [ ] Approve
- [ ] Reject

Tier 3 — More context required: For findings that need the user's judgment, append explicit questions to the same review file under ## Questions for you, each with the context needed to answer inline.

Step 4 — Report

## Improve System — YYYY-MM-DD
- Reviews processed: <n approved applied, n rejected dismissed>
- Auto-applied: <n> (see output/.improvement-log.md)
- Awaiting sign-off: <n> in output/review_YYYY-MM-DD.md
- Questions: <n>

If there were zero findings in every tier, say so and skip creating a review file. When run by a scheduled routine, commit and push to knowledge-base-updates as the routine prompt instructs — an ephemeral container discards uncommitted work. Otherwise do not commit unless the user asks.