Capture agent corrections and score recurring mistakes for rule promotion.
npx skills add https://github.com/qwwiwi/agentos-skills-public --skill learnings-qwwiwi
Please help me install this Agent Skill. Skill: learnings Source: https://github.com/qwwiwi/agentos-skills-public/tree/main/skills/learnings Command: npx skills add https://github.com/qwwiwi/agentos-skills-public --skill learnings-qwwiwi
Learnings System v2 helps you capture recurring errors and corrections, score them over time, and promote the best patterns into durable rules so the same failure modes don’t repeat.
Ask the agent to record a correction episode and then run scoring and linting to produce a learning report for promotion or archival.
High-intent search queries and answers about installing and using this skill.
Agent self-improvement turns repeated mistakes into enforceable rules by capturing explicit corrections into an append-only episode store. A composite scoring model using recency, frequency, and impact surfaces what matters most for durable system updates.
Error scoring calculates a composite score from recency, frequency, and impact to prioritize workflow audit findings. High-scoring learnings are proposed for owner-gated rule promotion to update system rules, while linting identifies hot, stale, or alert conditions.
Tracking error patterns across weeks requires an append-only episode store that records high-cost or recurring mistakes. This enables ongoing self-improvement by maintaining a durable learnings record accessible for linting and reporting across sessions.
Rule promotion is owner-gated, so you cannot automatically update rules. The system proposes turning high-scoring learnings into updated rules or CLAUDE.md, but the owner must explicitly review and approve the promotion to ensure quality management.
Linting checks for hot, stale, promote, and alert conditions within your episode store to identify maintenance needs. You can then generate reports from these linting results to review episodes for owner-gated rule promotion or archival.
Archive a learning episode when it becomes stale or fails to meet promotion thresholds during linting. Episodes with low composite scores from recency, frequency, and impact are better archived than promoted to durable rules.
yiwei79Captures session lessons as structured episodes with append-only governance and cross-linkages.
SPRAGEAudit CLAUDE.md files, generate quality reports, and apply approved updates.
hbvg234Capture failure state, diagnose root causes, and generate introspection reports for AI agent runs.
Chang-Shih-YungAggregate type checker, linter, and test results into a 0-10 composite score.
