self-improving-agent

• You frequently correct your AI agent — when you find yourself saying "No, that's wrong" or "Actually, use pnpm not npm," this skill captures those corrections so they aren't repeated.
• You're running a long-lived project — teams and solo developers who interact with AI agents across many sessions benefit from a growing, searchable record of project conventions and gotchas.
• You want to build reusable skills from solutions — when a non-obvious fix is discovered, the skill's extraction workflow turns it into a standalone, shareable skill.
• You use multi-agent workflows — OpenClaw users can broadcast learnings across sessions and spawn sub-agents to act on accumulated knowledge.

• Short, one-off tasks — if you're using an agent for a single throwaway script with no follow-up sessions, the overhead of logging and reviewing entries won't provide value.
• Fully automated pipelines with no human feedback loop — the skill's value compounds through user corrections and reviews; purely unattended runs won't generate useful signals.

Every captured event gets a typed, timestamped entry with priority, area tag, status, and metadata fields like Pattern-Key and See Also links. Three separate log files keep corrections, command failures, and feature requests organized and grep-friendly.

When a learning is broadly applicable, the skill provides a clear workflow to distill it into concise rules and write them to CLAUDE.md, AGENTS.md, .github/copilot-instructions.md, or OpenClaw workspace files (SOUL.md, TOOLS.md). This turns ephemeral session context into durable, contributor-visible knowledge.

Entries include a stable Pattern-Key and Recurrence-Count. When a pattern is logged 3+ times across 2+ distinct tasks within 30 days, the skill's promotion rules automatically flag it for elevation to system-prompt–level guidance, stopping repeat mistakes at the source.

Optional hook scripts (activator.sh, error-detector.sh) integrate with Claude Code and Codex CLI via UserPromptSubmit and PostToolUse hooks. This means error detection and learning prompts fire automatically without manual intervention (~50–100 token overhead per prompt).

Learnings that meet quality criteria (recurring, verified, non-obvious, broadly applicable) can be extracted into standalone, shareable skills using the provided extract-skill.sh helper. The workflow includes dry-run mode, quality gates, and status tracking on the original entry.

Works with Claude Code, Codex CLI, GitHub Copilot, and OpenClaw. Each agent has a documented activation method — hook-based for Claude Code/Codex, manual .github/copilot-instructions.md injection for Copilot, and workspace injection with inter-session messaging tools for OpenClaw.

An agent repeatedly tries npm install on a pnpm workspace. After the first correction, the skill logs a correction entry and the developer promotes it to CLAUDE.md as Package manager: pnpm. Subsequent sessions read this rule and never make the same mistake.

A team tracks .learnings/ in version control. When one developer's session hits a CI auth issue and logs it to ERRORS.md, every agent on the team picks up the documented fix and suggested resolution in future sessions.

After diagnosing a tricky TypeScript client regeneration issue that required three debugging sessions, the learning accumulates See Also links and a high Recurrence-Count. The developer runs extract-skill.sh api-client-regen to package the solution as a distributable skill.

Before starting a large refactor, an OpenClaw session uses the hook-triggered reminder to scan .learnings/ for pending high-priority entries in the backend area, applying all accumulated conventions before writing a single line of code.