Self Improving Agent.Bak

Captures learnings, errors, and corrections for continuous improvement. Use when: (1) a command fails, (2) user corrects you, (3) user requests a missing cap...

Security Scan

The skill's files, hooks, and scripts are consistent with its stated purpose (capturing and promoting learnings); it doesn't request credentials or perform network installs, but you should review and constrain hook activation and learning contents before enabling globally.

Purpose & Capability

Name/description (capture learnings, errors, promotions) match the included artifacts: SKILL.md, hook handlers that inject a bootstrap reminder, and scripts to activate/error-detect and extract skills. No extraneous credentials, binaries, or network installs are requested.

Instruction Scope

Runtime instructions and scripts focus on writing/readings under a .learnings folder, injecting a virtual reminder at agent bootstrap, and inspecting CLAUDE_TOOL_OUTPUT for errors. They do not instruct reading unrelated system files or exfiltrating data, but if enabled at a user/global level the hooks will run frequently and could capture whatever is written into .learnings (so avoid logging secrets).

Install Mechanism

No install spec or external downloads. All code is included in the package (hooks and shell scripts); nothing is fetched from external URLs and no archives are extracted.

Credentials

The skill declares no required environment variables or credentials. The error-detector script reads CLAUDE_TOOL_OUTPUT (expected for a PostToolUse hook) which aligns with its purpose. No unrelated secrets or keys are requested.

Persistence & Privilege

always is false and the skill is user-invocable. Hook files can be copied/enabled in OpenClaw, but the package does not demand forced or system-wide persistence nor modify other skills' configs. Note: enabling hooks in user-level settings will make them run for that user/session.

Assessment

This skill appears to do what it says: inject lightweight reminders, detect command errors from CLAUDE_TOOL_OUTPUT, and help scaffold extracted skills. Before installing/activating: 1) Inspect the scripts (activator.sh, error-detector.sh, extract-skill.sh) and hook handler to ensure they match your expectations. 2) Prefer project-level hook configuration (not user/global) to limit how often hooks run. 3) Do not log secrets into .learnings/ (these files may be stored in your workspace and could be visible to other tools). 4) Keep script permissions conservative (chmod +x only after review). 5) If you plan to run extract-skill.sh, note it will create files under a relative ./skills directory (the script prohibits absolute/.. paths). If you want extra assurance, enable the hook in a sandbox workspace first and confirm behavior.

Self-Improvement Skill

Log learnings to .learnings/ for continuous improvement. Promote broadly applicable entries to workspace files.

Setup

mkdir -p ~/.openclaw/workspace/.learnings

Create three log files: LEARNINGS.md · ERRORS.md · FEATURE_REQUESTS.md

Routing Table

Situation	File	Category
Command/operation fails	ERRORS.md	—
User corrects you	LEARNINGS.md	correction
Missing capability requested	FEATURE_REQUESTS.md	—
Knowledge was outdated	LEARNINGS.md	knowledge_gap
Found better approach	LEARNINGS.md	best_practice
Recurring pattern (simplify-and-harden)	LEARNINGS.md	Source: simplify-and-harden + Pattern-Key

Promotion Targets

Learning Type	Promote To
Behavioral patterns	SOUL.md
Workflow improvements	AGENTS.md
Tool gotchas	TOOLS.md
Project facts / conventions	CLAUDE.md

Promote when: applies across multiple files, prevents recurring mistakes, or any contributor should know it.

Log Formats

LEARNINGS.md entry

## [LRN-YYYYMMDD-XXX] category
**Logged**: ISO-8601  **Priority**: low|medium|high|critical  **Status**: pending  **Area**: frontend|backend|infra|tests|docs|config

### Summary
One-line description

### Details
What happened, what was wrong, what's correct

### Suggested Action
Specific fix or improvement

### Metadata
- Source: conversation|error|user_feedback
- Related Files:
- Tags:
- See Also: (if related entry exists)
- Pattern-Key: (optional, for recurring patterns)
- Recurrence-Count: 1
---

ERRORS.md entry

## [ERR-YYYYMMDD-XXX] command_name
**Logged**: ISO-8601  **Priority**: high  **Status**: pending  **Area**: …

### Summary / Error / Context / Suggested Fix
### Metadata
- Reproducible: yes|no|unknown
- See Also:
---

FEATURE_REQUESTS.md entry

## [FEAT-YYYYMMDD-XXX] capability_name
**Logged**: ISO-8601  **Priority**: medium  **Status**: pending  **Area**: …

### Requested Capability / User Context / Complexity Estimate / Suggested Implementation
### Metadata
- Frequency: first_time|recurring
---

Resolving Entries

Change **Status**: pending → resolved | in_progress | wont_fix | promoted, then add:

### Resolution
- **Resolved**: ISO-8601
- **Notes**: what was done

Recurring Pattern Detection

1. Search first: grep -r "keyword" .learnings/
2. If similar entry exists: increment Recurrence-Count, update Last-Seen, add See Also
3. Promote to system prompt files when: Recurrence-Count >= 3 + seen in 2+ tasks + within 30 days

Periodic Review

grep -h "Status\*\*: pending" .learnings/*.md | wc -l          # count pending
grep -B5 "Priority\*\*: high" .learnings/*.md | grep "^## \["  # list high-priority

Review before major tasks, after completing features, or weekly during active development.

Detection Triggers

• Correction: "No, that's wrong…" / "Actually…" → LEARNINGS.md correction
• Feature request: "Can you also…" / "I wish you could…" → FEATURE_REQUESTS.md
• Knowledge gap: user provides info you didn't know → LEARNINGS.md knowledge_gap
• Error: non-zero exit / exception / timeout → ERRORS.md

Skill Extraction

Extract a learning into a reusable skill when: recurring (2+ See Also links) + resolved + non-obvious + broadly applicable.

./skills/self-improvement/scripts/extract-skill.sh skill-name

Or manually: create skills/<name>/SKILL.md with YAML frontmatter (name, description).

Priority / Area Reference

Priority: critical (blocks core / data loss) · high (common workflow) · medium (workaround exists) · low (edge case)

Area: frontend · backend · infra · tests · docs · config