Self Improving Agent

Capture errors, corrections, and recurring patterns into structured `.learnings/` logs, then promote durable guidance into workspace memory files. Use when c...

Security Scan

Files, scripts, and hook handlers match the stated purpose of capturing and promoting development learnings; nothing requests unrelated credentials or performs covert networking or writes outside the workspace without user action.

Purpose & Capability

Name/description align with included files: reminder hook, activator, error detector, extraction helper, and templates. The skill does not request unrelated credentials or binaries and the provided handlers/scripts implement the stated capture-and-promote workflow.

Instruction Scope

SKILL.md and references instruct installing hooks and running scripts that output reminders and optionally create skill scaffolds. The error-detector reads the CLAUDE_TOOL_OUTPUT env var (expected for a PostToolUse hook). The extract-skill.sh script does create files under a relative ./skills/ directory when run (it has safeguards against absolute or ../ paths). Be aware hooks can be configured at user-level (~/.claude) which makes the activator run more broadly — this is a functional choice but increases execution surface.

Install Mechanism

No remote downloads or install steps; instruction-only skill with bundled scripts and hook handlers. There is no installer that fetches remote code or writes binaries to unexpected locations.

Credentials

No credentials or sensitive environment variables are requested. The only environment usage of note is error-detector.sh reading CLAUDE_TOOL_OUTPUT (a platform-provided variable containing tool output) and check_env.sh inspecting local paths — both are proportionate to detecting and logging errors.

Persistence & Privilege

always is false and the hook handler injects a virtual bootstrap file (no persistent modification). However, enabling hooks at user-level or project-level will cause the activator/error-detector scripts to run automatically on configured events; enabling globally increases runtime frequency and blast radius, so prefer project-scoped configuration if you want limited scope.

Assessment

This skill appears to do what it claims, but review and control how you enable it: 1) Inspect the scripts (activator.sh, error-detector.sh, extract-skill.sh) before enabling hooks to confirm behavior. 2) Prefer project-level hook configuration (not global ~/.claude) if you want to limit automatic executions. 3) Note extract-skill.sh will create files under ./skills/ when run (it prevents absolute/../ paths). 4) error-detector reads CLAUDE_TOOL_OUTPUT (it may contain command output — avoid logging secrets into .learnings/). 5) Run scripts/check_env.sh manually to verify paths and permissions in your workspace before enabling automated hooks.

Use this skill to turn execution feedback into reusable operational knowledge.

Quick workflow

1. Detect signal: error, correction, capability gap, or repeated workaround.
2. Log to .learnings/ with the right entry type.
3. Resolve or promote high-value patterns into durable workspace files.
4. Reuse scripts/hooks to keep capture consistent.

Use bundled resources

• Entry examples and formats: references/examples.md
• Hook setup for reminders/error detection: references/hooks-setup.md
• OpenClaw workspace integration: references/openclaw-integration.md
• Reminder hook script: scripts/activator.sh
• Command-error detector: scripts/error-detector.sh
• Skill extraction scaffold: scripts/extract-skill.sh
• Environment checks: scripts/check_env.sh
• Log templates: assets/LEARNINGS.md, assets/SKILL-TEMPLATE.md

Log targets

• .learnings/LEARNINGS.md: corrections, knowledge gaps, best practices
• .learnings/ERRORS.md: command/tool/runtime failures
• .learnings/FEATURE_REQUESTS.md: requested capabilities not yet supported

Promotion rules

Promote broadly reusable learnings out of .learnings/:

• behavior/style -> SOUL.md
• workflow/orchestration -> AGENTS.md
• tool constraints/gotchas -> TOOLS.md

Update original entries with status transitions (resolved, promoted, wont_fix) and references.

Commands

# Verify scripts and learnings directory setup
bash workspace/skills/self-improving-agent/scripts/check_env.sh

# Dry-run extraction of a new skill from a recurring pattern
bash workspace/skills/self-improving-agent/scripts/extract-skill.sh my-pattern --dry-run