My MBS Self-Improvement Log

What It Solves

AI coding agents often hit command failures, user corrections, stale APIs, or project conventions that are only visible inside one conversation. Without a durable record, the same mistakes can resurface in later sessions. My MBS turns those process-level signals into structured Markdown logs so errors, corrections, feature requests, and project learnings can be reviewed, tracked, and promoted.

How It Works

On first use, it initializes .learnings/ with ERRORS.md, LEARNINGS.md, and FEATURE_REQUESTS.md at the project or workspace root. When a command exits non-zero, a stack trace appears, a user corrects the agent, an external tool fails, or knowledge is outdated, the skill records the event as an entry with an ID, Status, Priority, and Area. For recurring patterns, it deduplicates by Pattern-Key, updates Recurrence-Count and Last-Seen, and distills high-confidence signals into shorter prevention rules.

Learnings that prove useful across tasks are promoted to CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md; in OpenClaw workspaces they may also go into SOUL.md or TOOLS.md. The skill does not automatically patch code. It maintains an auditable memory pipeline: log the event, review recurring signals, promote stable rules, then mark the original entry as resolved or promoted. Optional hooks can remind the agent to evaluate the log after tasks, but only when the project explicitly wants that extra workflow.

Where It Fits

It suits long-lived engineering workspaces where agent rules, tool gotchas, and project conventions need to persist across sessions. Avoid recording secrets, tokens, environment variables, or full config files by default. Cross-session sharing should only happen in trusted environments when the user explicitly asks for it. Logs can remain local for personal use or be tracked in the repo as shared team knowledge.