Self Improving Agent

Create .learnings/LEARNINGS.md, .learnings/ERRORS.md, and .learnings/FEATURE_REQUESTS.md in the project or OpenClaw workspace. If the files already exist, keep the current entries and append new learnings instead of overwriting them.

When a command fails, a user corrects the agent, an API behaves differently than expected, or a better recurring approach is discovered, add a short structured entry with priority, area, status, summary, suggested action, and related files.

Before starting a refactor, migration, deployment, or multi-step automation, search .learnings/ for the relevant area and apply pending high-priority guidance. This is the core self improving skill loop: capture, review, apply, and refine.

When the same pattern appears across tasks, distill it into a short rule in AGENTS.md, CLAUDE.md, .github/copilot-instructions.md, SOUL.md, or TOOLS.md. Mark the original learning as promoted so the log stays useful.

• You want a self improving agent skill that records corrections, errors, and best practices as structured project knowledge.
• You run OpenClaw across long-lived repos and want repeatable workspace memory instead of relying on chat history.
• You use Claude Code, Codex, GitHub Copilot, or another coding agent and need a self improving skill that works outside one session.
• You install or publish skills through ClawHub and want a practical loop for turning recurring fixes into reusable skill guidance.
• You often correct the same package manager, API, command, architecture, or style mistake and want future agents to catch it earlier.

• You only need a one-off script or throwaway agent session where no future learning will be reused.
• You want fully autonomous self-modification with no human review. This skill logs and promotes learnings, but important rules should still be reviewed.
• Your workflow would require storing secrets, private keys, full transcripts, or sensitive command output in logs. Use redacted summaries instead.
• You need a generic memory database or vector search product rather than a markdown-based agent skill.

The skill creates and maintains .learnings/LEARNINGS.md, .learnings/ERRORS.md, and .learnings/FEATURE_REQUESTS.md. Each entry can include priority, status, area, summary, suggested action, related files, pattern keys, recurrence counts, and see-also links so future agents can search and reuse the knowledge quickly.

When a correction becomes broadly useful, the workflow distills it into a concise rule for CLAUDE.md, AGENTS.md, .github/copilot-instructions.md, or OpenClaw workspace files like SOUL.md and TOOLS.md. This turns a temporary chat correction into durable project guidance.

Stable Pattern-Key values and Recurrence-Count metadata help identify repeated problems. When the same issue appears across tasks, the skill gives you a clear path to promote the prevention rule instead of logging another isolated incident.

Optional hook workflows can remind Claude Code or Codex to evaluate new prompts, command failures, and user corrections. The agent does not need to remember every past conversation because the important patterns are written into readable project files.

OpenClaw users can keep .learnings/ inside the workspace, while Claude Code, Codex, GitHub Copilot, and similar agents can consume the same markdown memory files. That makes the self improving skill OpenClaw-friendly without locking the learning loop to one agent runtime.

Hard-won solutions can be packaged into reusable skills when they are recurring, verified, non-obvious, and broadly applicable. This makes the page relevant for users searching for a self improving skill on ClawHub as well as teams building their own skill library.

When an agent keeps using the wrong package manager, command, API pattern, import path, or deployment step, log the correction once and promote it when it recurs. Future sessions can read the rule before touching the same area.

Use Self Improving Agent inside an OpenClaw workspace to capture local conventions, tool gotchas, and workflow patterns. Before a refactor or automation task, the agent can review relevant learnings and apply them immediately.

Claude Code sessions can use the logs to remember recurring project facts such as test commands, migration rules, API response formats, UI constraints, and review preferences. The result is a self improving skill Claude users can apply without rewriting instructions every time.

If a recurring solution becomes valuable outside one project, use the extraction workflow to turn it into a standalone skill. This helps teams move from one-off fixes to reusable ClawHub-ready agent skills.