From core-aiws
Analyze accumulated user signals and propose improvements to CLAUDE.md, agents, skills, and hooks
How this skill is triggered — by the user, by Claude, or both
Slash command
/core-aiws:improveThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
This is the shared `/aiws-improve` surface owned by `core-aiws`.
This is the shared /aiws-improve surface owned by core-aiws.
Gathers accumulated signals from multiple sources, synthesizes patterns, then runs the unified Self-Improvement Protocol in batch mode.
Scope: This skill is responsible only for evidence gathering and synthesis (Phases 1-3). All decision rules for prompt, skill, protocol, and workflow improvement live in the protocol — do not duplicate them here. Shared-memory refresh is not owned by this skill.
Read all available sources (skip any that don't exist):
${CLAUDE_PLUGIN_DATA}/improve/observations.jsonl — find the most recent improve_run entry as cutoff, filter to newer entries${CLAUDE_PLUGIN_DATA}/project-memory/current/YYYY-MM-DD.md${CLAUDE_PLUGIN_DATA}/registry/plugins/*.jsonPresent evidence summary:
## Evidence Summary (since last /aiws-improve run)
**Observations** (from hook signals):
| Signal Type | Count |
|---------------|-------|
| correction | N |
| frustration | N |
| give_up | N |
| positive | N |
**Other sources**: N daily log entries, N session histories reviewed
- Unique sessions: N
- Unique projects: N
- Date range: YYYY-MM-DD to YYYY-MM-DD
If no evidence exists from any source, report "No new signals to analyze" and stop.
For each high-severity observation (correction, frustration, give_up):
transcript pathPresent findings:
### Finding: <obs_id> (<type>, <date>)
**User said**: "<message excerpt>"
**Context**: <what Claude was doing>
**Root cause**: <category> - <specific explanation>
**Target**: <file path> : <section/line>
Group findings by root cause across sessions:
Present:
### Pattern: <descriptive name>
- Sessions: <list of session dates>
- Root cause: <category>
- Evidence: "<quote 1>", "<quote 2>"
- Target file: <path>
- Confidence: HIGH/MEDIUM (see protocol rules)
Follow the Self-Improvement Protocol in batch mode with the synthesized findings from Phase 3 as input. Start from Step 3 (Categorize and Decide) — Steps 1-2 are skipped in batch mode. Use the synthesized patterns from Phase 3 as input to Step 3's categorization; formal Learning Entry Format is applied in Step 4.2.
Do not treat /aiws-improve as the routine shared-memory consolidation trigger. Shared-memory candidate capture happens during end-of-task auto-capture, and shared-memory refresh is handled automatically by the host-side shared-memory bridge.
After protocol completion:
Append an improve_run marker to ${CLAUDE_PLUGIN_DATA}/improve/observations.jsonl:
{"id":"imp_<8-char-hex>","ts":"<ISO timestamp>","type":"improve_run","severity":"info","message":"Processed observations up to <latest_obs_id>"}
For each applied change, append a verification entry:
{"id":"verify_<8-char-hex>","ts":"<ISO timestamp>","type":"improvement_applied","severity":"info","message":"Applied: <brief description>. Monitor for recurrence."}
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.
npx claudepluginhub sashakang/ai-workspace --plugin core-aiws