self_improving_agent
v1.0.0Captures learnings, errors, and corrections to enable continuous improvement. Use when: (1) A command or operation fails unexpectedly, (2) User corrects Clau...
⭐ 0· 191·0 current·0 all-time
MIT-0
Download zip
LicenseMIT-0 · Free to use, modify, and redistribute. No attribution required.
Security Scan
OpenClaw
Benign
high confidencePurpose & Capability
The skill is designed to capture learnings/errors and promote them into workspace files or new skills. The included scripts (activator, error-detector, extract-skill) and OpenClaw/handler hooks all relate directly to that purpose. Minor inconsistency: packaged _meta.json and the registry metadata show different ownerId/version timestamps, which could indicate the bundled files came from a different upstream snapshot than the registry entry—worth verifying the source before install.
Instruction Scope
SKILL.md and hooks instruct the agent to create and update .learnings/ files, inject reminder content into sessions, and optionally promote entries into workspace-level files (CLAUDE.md, AGENTS.md, SOUL.md). The scripts only read platform-provided environment (CLAUDE_TOOL_OUTPUT in error-detector) and operate on local workspace paths. This is within the skill's stated scope, but the promotion/injection behavior means the skill can change the agent's context (prompt injection effect) if enabled globally—this is expected for this class of tool but important to understand.
Install Mechanism
There is no automated install spec; installation is manual (clawdhub install or git clone from GitHub). All included code is in the package (no remote downloads or network installs during runtime). The git URL referenced is a normal GitHub repository. No high-risk download URLs or archive extraction steps are present.
Credentials
The skill does not request credentials or environment variables. The error-detector script reads a platform-provided variable (CLAUDE_TOOL_OUTPUT) to detect failures—this is proportional to detecting tool errors. Scripts write to project/user workspace (e.g., ~/.openclaw/workspace/.learnings or ./ .learnings) which is consistent with logging behavior; extract-skill.sh validates output paths to avoid writing outside the workspace.
Persistence & Privilege
always:false and no primary credential are good. However, the documentation encourages copying hooks to user-level locations (~/.openclaw/hooks) and enabling them in user/global settings; if a user enables those hooks globally, the reminder injection will affect every session for that user. This is intentional for the skill but increases its reach—enable at project scope if you want to limit impact.
Assessment
This skill appears to do what it says: add lightweight reminders and local logging of errors/learnings and offer a helper to scaffold new skills from those learnings. Before installing or enabling hooks: (1) verify the repository/source you clone matches the registry entry (the included _meta.json shows different metadata—confirm origin), (2) review the scripts (they only emit text and manipulate files in your workspace, but inspect extract-skill.sh before running to confirm it meets your expectations), (3) prefer project-level hook configuration rather than user-global if you want to avoid affecting all sessions, and (4) be aware that promoted files (CLAUDE.md, SOUL.md, etc.) are injected into agent context and will change agent behavior—treat that as a deliberate prompt-injection mechanism and review promoted content before making it global.Like a lobster shell, security has layers — review code before you run it.
latestvk978fs4ejncb6c080anfqv66258399vf
License
MIT-0
Free to use, modify, and redistribute. No attribution required.