Murphe Self-Improvement

Problem

When AI agents make mistakes during coding, debugging, or tool use, the lesson often stays inside one session. The next time the agent hits the same API behavior, package-manager constraint, or user correction, it may repeat the error. This skill turns short-lived conversational memory into reviewable Markdown logs so errors, corrections, feature requests, and best practices can be processed by later agents, developers, or review workflows.

How It Works

It centers on a .learnings/ directory. Command failures go to ERRORS.md; user corrections and knowledge gaps go to LEARNINGS.md; missing capabilities go to FEATURE_REQUESTS.md. Each entry carries metadata such as ID, Status, Priority, Area, and See Also, which makes filtering by frontend, backend, or infra practical. When an entry recurs or proves broadly useful, it is distilled into concise rules and promoted into context files such as CLAUDE.md, AGENTS.md, and .github/copilot-instructions.md; OpenClaw workspaces can also use SOUL.md and TOOLS.md. For recurring patterns, the skill supports Pattern-Key dedupe, Recurrence-Count tracking, and promotion after recurrence, cross-task, and 30-day conditions, producing prevention rules rather than long incident write-ups.

Boundaries

This skill fits continuously evolving codebases and multi-agent workflows, especially environments that already use AGENTS.md or CLAUDE.md. It is not an automatic fixer; its main outputs are structured logs and promotable project memory. Teams should decide whether to track or ignore .learnings/ in Git. The more specific and earlier the entries are written, the more reusable they become; overly broad notes can add context noise instead of durable guidance.