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...
⭐ 5· 10.1k·221 current·227 all-time
MIT-0
Download zip
LicenseMIT-0 · Free to use, modify, and redistribute. No attribution required.
Security Scan
OpenClaw
Benign
high confidencePurpose & Capability
Name/description (self-improvement / log learnings) align with the included files: reminder hook, activator and error-detector scripts, learning templates, and an extract-skill helper. The code and docs are consistent with logging and scaffolding skills; no unrelated cloud credentials, binaries, or surprising external services are required.
Instruction Scope
SKILL.md and hook code limit behavior to: injecting a virtual reminder at bootstrap, printing a short activator reminder after prompts, detecting simple error patterns from tool output, and scaffolding new skill folders from learnings. There are no instructions to read or transmit arbitrary system secrets, to crawl unrelated files, or to call external endpoints. One note: the error-detector reads the CLAUDE_TOOL_OUTPUT environment variable (agent-provided tool output) to decide whether to emit a reminder.
Install Mechanism
This is instruction-only (no automated installer). Manual install suggestions reference git clone from a GitHub repo (public host). No downloads from obscure URLs or archive extraction are present. The included scripts are small, local shell scripts and a JS/TS hook handler for OpenClaw.
Credentials
The skill declares no required environment variables or credentials. The error-detector script does read CLAUDE_TOOL_OUTPUT (an agent-provided environment variable) and the hook interacts with OpenClaw event objects — both are reasonable for an agent hook. Keep in mind logs and captured tool outputs could contain sensitive data (commands, error traces, paths); the skill does not declare or require secrets, but enabling PostToolUse hooks may surface tool outputs that include secrets unless you configure filters/scrubbing.
Persistence & Privilege
always:false (not force-included). Hooks are opt-in: you must copy/enable them (openclaw hooks enable ...) or add them to agent settings. The extract-skill script writes under a relative ./skills directory and explicitly disallows absolute paths and '..' traversal. Handlers inject a virtual bootstrap file at runtime only (they modify event context, not global agent config).
Assessment
This skill appears to be what it says: a local helper to capture and promote learnings. Before enabling hooks or PostToolUse scripts, review and test them in a safe environment: the error-detector reads CLAUDE_TOOL_OUTPUT and may echo command/tool output (which can include passwords, tokens, or other secrets), so avoid running hooks that capture tool output for commands that handle secrets. If you enable the OpenClaw hook, note it injects a virtual reminder at bootstrap and the activator prints a short reminder on prompts — harmless but potentially noisy. The extract-skill script creates files in your workspace; review the created SKILL.md before committing it. If you have privacy concerns, keep the .learnings folder in a private workspace or add scrubbing/filtering to the hooks.Like a lobster shell, security has layers — review code before you run it.
latestvk976j8wnf31tn03dcejbrak0r1824pkm
License
MIT-0
Free to use, modify, and redistribute. No attribution required.