AI Agent Hub
Back to skills
Self-Improvement Log icon

Self-Improvement Log

AI Agent Updated 2026.08.30

Paste the following prompt into your AI chat to install this skill:

Follow https://skillhub.cn/install/skillhub.md to install @user_2908d46b/skill-js.

About this skill

Problem

AI agents often encounter one-off failures, user corrections, stale knowledge, or tool quirks that get lost once context moves on. Later sessions repeat the same mistakes, and teams lack a lightweight way to turn facts such as “this project uses pnpm” or “Git push needs auth configured” into searchable, persistent guidance. This leaves recurring failures harder to debug and less visible to humans.

Workflow

The skill uses Markdown files under .learnings/ as external memory: errors go to ERRORS.md, corrections, knowledge gaps, and best practices go to LEARNINGS.md, and capability requests go to FEATURE_REQUESTS.md. Each entry has a stable ID such as LRN-20250115-001, plus Status, priority, Area, and See Also links. The record is not just an archive; it feeds later fixes, reviews, and skill extraction. When a learning applies across files or contributors, it is distilled into CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md; in OpenClaw, it may also move to SOUL.md or TOOLS.md. The workflow also asks the agent to search existing entries first, merge duplicates, and escalate recurring issues into systemic fixes.

Boundaries

It relies on the agent following the rules or on hook-based reminders; it does not automatically patch errors or publish skills. If sensitive information must stay out of the repo, keep .learnings/ local, or track only templates while ignoring entries.

Use Cases

  • Log failed commands with error output and repro steps to `ERRORS.md` for later debugging sessions.
  • After a user corrects the agent, log the correction category, source, and scope to `LEARNINGS.md`.
  • When a tool quirk recurs, link related entries and distill a prevention rule into `AGENTS.md` or `CLAUDE.md`.
  • Before converting a reusable learning into a skill, verify status, `See Also` links, and draft `SKILL.md`.

Best For

  • Engineers using Claude Code for daily development who need to turn errors and corrections into project rules.
  • Engineers maintaining multi-agent workflows who need a consistent way to capture learnings, review recurrence, and extract skills.
  • Developers building OpenClaw assistants who need to move high-frequency behavioral rules into workspace memory files.
  • Platform maintainers sharing agent experience with a team who want to commit templates while keeping sensitive entries local.