Self-Improving Meta Skill

Log infrastructure learnings, meta issues, and feature requests to markdown files for continuous improvement of the agent system itself. Captures prompt drift, rule conflicts, skill gaps, hook failures, context bloat, and instruction ambiguity. Important learnings may be promoted into the files they govern — prompt files, hook code, rule files, skill templates, and memory policies — after explicit human review.

This is the skill that logs infrastructure problems other skills depend on. It does not apply infrastructure edits by default. Propose patches; apply them only when the user asks in this session.

First-Use Initialisation

Before logging anything, ensure the .learnings/ directory and files exist in the project or workspace root. If any are missing, create them:

mkdir -p .learnings
[ -f .learnings/LEARNINGS.md ] || printf "# Meta Learnings\n\nPrompt drift, rule conflicts, skill gaps, hook failures, context bloat, and instruction ambiguity.\n\n**Categories**: prompt_drift | rule_conflict | skill_gap | hook_failure | context_bloat | instruction_ambiguity\n**Areas**: agent_config | skill_authoring | hook_scripts | prompt_files | rule_files | memory_management | extension_api\n\n---\n" > .learnings/LEARNINGS.md
[ -f .learnings/META_ISSUES.md ] || printf "# Meta Issues Log\n\nInfrastructure failures: hook crashes, skill activation problems, prompt file errors, memory corruption, extension breakage.\n\n---\n" > .learnings/META_ISSUES.md
[ -f .learnings/FEATURE_REQUESTS.md ] || printf "# Feature Requests\n\nCapabilities needed for agent infrastructure, skill authoring, hook development, and prompt management.\n\n---\n" > .learnings/FEATURE_REQUESTS.md

Never overwrite existing files. This is a no-op if .learnings/ is already initialised.

Do not log secrets, tokens, private keys, or environment variables. Prefer short summaries or redacted excerpts over raw file contents.

Use a manual-first workflow by default. If you want reminders, use the opt-in hook workflow described in Hook Integration.

Safety Boundaries

• Default is log-only. Do not edit AGENTS.md, SOUL.md, TOOLS.md, MEMORY.md, hooks, rules, or skill files unless the user explicitly asked for that edit in the current session.
• Prefer proposing a minimal patch and rationale. Never silent self-modification.
• Treat hook output as sensitive; avoid logging raw command output or full transcripts.
• Keep fixes scoped to the identified issue; avoid broad refactors during incident response.

Quick Reference

OpenClaw Setup (Recommended)

OpenClaw is the primary platform for this skill. It uses workspace-based prompt injection with automatic skill loading.

Installation

Prefer the registry install (network to ClawHub):

clawdhub install self-improving-meta

Manual clone is optional network access to GitHub. Use only if you trust that repository:

git clone https://github.com/jose-compu/self-improving-meta.git ~/.openclaw/skills/self-improving-meta

Workspace Structure

OpenClaw injects these files into every session:

~/.openclaw/workspace/
├── AGENTS.md          # Multi-agent workflows, delegation patterns
├── SOUL.md            # Behavioral guidelines, personality, principles
├── TOOLS.md           # Tool capabilities, integration gotchas
├── MEMORY.md          # Long-term memory (main session only)
├── memory/            # Daily memory files
│   └── YYYY-MM-DD.md
└── .learnings/        # This skill's log files
    ├── LEARNINGS.md
    ├── META_ISSUES.md
    └── FEATURE_REQUESTS.md

Create Learning Files

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

Then create the log files (or copy from assets/):

• LEARNINGS.md — prompt drift, rule conflicts, skill gaps, context bloat, instruction ambiguity
• META_ISSUES.md — hook crashes, skill activation failures, prompt file errors, memory corruption
• FEATURE_REQUESTS.md — infrastructure capabilities, tooling, automation

Promotion Targets

When meta-learnings prove broadly applicable, promote them to the files they govern with explicit user approval:

Optional: Enable Hook

Two different hook modes exist. Do not mix their semantics:

1. Claude Code / Codex (recommended if you want automation): matcher-gated, project .claude/settings.json / .codex/settings.json. Fires only on explicit meta signals.
2. OpenClaw agent:bootstrap (optional): not matcher-gated. If enabled, it injects a log-only reminder at main-session start. It does not edit files. Skip it unless you want that reminder.

Do not install into ~/.openclaw/hooks/ (user-global). If you enable the OpenClaw hook, keep it in this workspace only:

mkdir -p .openclaw/hooks
cp -r hooks/openclaw .openclaw/hooks/self-improving-meta

See references/openclaw-integration.md for complete details.

Generic Setup (Other Agents)

For Claude Code, Codex, Copilot, or other agents, create .learnings/ in the project or workspace root:

mkdir -p .learnings

Create the files inline using the headers shown above.

Add reference to agent files

Add to AGENTS.md, CLAUDE.md, or .github/copilot-instructions.md:

Self-Improving Meta Workflow

When agent infrastructure issues are discovered:

1. Log to .learnings/META_ISSUES.md, LEARNINGS.md, or FEATURE_REQUESTS.md
2. After explicit user approval of a reviewed diff, promote broadly applicable learnings to:
   • SOUL.md — behavioral corrections
   • AGENTS.md — workflow and delegation improvements
   • TOOLS.md — tool integration fixes
   • MEMORY.md — memory management policies
   • Rule files — .cursor/rules/, AGENTS.md rules
   • Hook code — propose a handler patch and apply after approval

Logging Format

Learning Entry [LRN-YYYYMMDD-XXX]

Append to .learnings/LEARNINGS.md:

## [LRN-YYYYMMDD-XXX] category

**Logged**: ISO-8601 timestamp
**Priority**: low | medium | high | critical
**Status**: pending
**Area**: agent_config | skill_authoring | hook_scripts | prompt_files | rule_files | memory_management | extension_api

### Summary
One-line description of the infrastructure insight

### Details
Full context: what prompt file instruction was ambiguous, which rules conflict,
why context is bloated, how memory drifted. Include the relevant file paths and
line numbers. Quote the problematic instruction verbatim.

### Suggested Action
Specific fix: rewrite the instruction, compress the file, consolidate the rules,
update the hook code, prune the memory entries.

### Metadata
- Source: agent_behavior_observation | hook_output_inspection | context_window_analysis | learnings_review | prompt_audit
- Affected Files: AGENTS.md, SOUL.md, etc.
- Tags: tag1, tag2
- See Also: LRN-20250110-001 (if related to existing entry)
- Pattern-Key: rule_conflict.package_manager (optional)
- Recurrence-Count: 1 (optional)
- First-Seen: 2025-01-15 (optional)
- Last-Seen: 2025-01-15 (optional)

---

Categories for learnings:

Meta Issue Entry [META-YYYYMMDD-XXX]

Append to .learnings/META_ISSUES.md:

## [META-YYYYMMDD-XXX] issue_type

**Logged**: ISO-8601 timestamp
**Priority**: high
**Status**: pending
**Area**: agent_config | skill_authoring | hook_scripts | prompt_files | rule_files | memory_management | extension_api

### Summary
Brief description of the infrastructure failure

### Error Output
\`\`\`
Actual error message, hook output, or observed misbehavior
\`\`\`

### Root Cause
What in the infrastructure caused this failure. Include the affected file and section.

### Suggested Fix
How to resolve: update hook code, fix skill frontmatter, correct prompt file, prune memory.

### Metadata
- Source: hook_output_inspection | skill_activation_failure | prompt_injection_error | memory_corruption | extension_api_breakage
- Affected Files: path/to/file
- Reproducible: yes | no | unknown
- See Also: META-20250110-001 (if recurring)

---

Feature Request Entry [FEAT-YYYYMMDD-XXX]

Append to .learnings/FEATURE_REQUESTS.md:

## [FEAT-YYYYMMDD-XXX] capability_name

**Logged**: ISO-8601 timestamp
**Priority**: medium
**Status**: pending
**Area**: agent_config | skill_authoring | hook_scripts | prompt_files | rule_files | memory_management | extension_api

### Requested Capability
What infrastructure tool, automation, or capability is needed

### User Context
Why it's needed, what workflow it improves, what infrastructure problem it solves

### Complexity Estimate
simple | medium | complex

### Suggested Implementation
How this could be built: hook script, skill template, audit script, linting tool

### Metadata
- Frequency: first_time | recurring
- Related Features: existing_tool_or_feature

---

ID Generation

Format: TYPE-YYYYMMDD-XXX

• TYPE: LRN (learning), META (meta/infrastructure issue), FEAT (feature request)
• YYYYMMDD: Current date
• XXX: Sequential number or random 3 chars (e.g., 001, A7B)

Examples: LRN-20250415-001, META-20250415-A3F, FEAT-20250415-002

Resolving Entries

When an issue is fixed, update the entry:

1. Change **Status**: pending → **Status**: resolved
2. Add resolution block after Metadata:

### Resolution
- **Resolved**: 2025-01-16T09:00:00Z
- **Fix Applied To**: AGENTS.md / SOUL.md / hook code / skill template
- **Notes**: Rewrote delegation instruction / compressed SOUL.md / fixed hook path

Other status values:

• in_progress — Actively investigating or fixing
• wont_fix — Decided not to address (add reason in Resolution notes)
• promoted — Elevated after explicit user approval of a reviewed diff
• promoted_to_skill — Extracted as a reusable meta-skill

Promoting to Project Memory

Meta-learnings are special: they can affect shared infrastructure. When you improve a prompt file, that improvement affects future sessions. When you fix a hook, that fix can propagate to all bootstraps. Use review gates.

When to Promote

• Same misinterpretation occurs across 2+ sessions
• Rule conflict is confirmed (not a one-off edge case)
• Context savings exceed 500 tokens after compression
• Hook failure affects multiple sessions or agents
• Skill gap has 3+ related learnings ready for extraction

Promotion Targets

How to Promote

1. Distill the learning into a concrete fix
2. Prepare a minimal patch for the affected file
3. Test in a fresh session to verify the fix works
4. Show the reviewed diff and apply only after explicit user approval, then update original entry:
   • Change **Status**: pending → **Status**: promoted
   • Add **Promoted**: AGENTS.md (delegation section rewrite) (or equivalent)

Recurring Pattern Detection

If logging something similar to an existing entry:

1. Search first: grep -r "keyword" .learnings/
2. Link entries: Add **See Also**: LRN-20250110-001 in Metadata
3. Bump priority if issue keeps recurring
4. Consider systemic fix: Recurring meta issues often indicate:
   • Instruction needs rewriting (→ propose a prompt-file patch; apply after approval)
   • Rule belongs in a different file (→ relocate)
   • Hook needs hardening (→ add validation/output checks)
   • Skill template is incomplete (→ update template)

Simplify & Harden Feed

Ingest recurring infrastructure patterns from simplify-and-harden into prompt file fixes and hook improvements.

1. For each candidate, use pattern_key as the dedupe key.
2. Search .learnings/LEARNINGS.md for existing entry: grep -n "Pattern-Key: <key>" .learnings/LEARNINGS.md
3. If found: increment Recurrence-Count, update Last-Seen, add See Also links.
4. If not found: create new LRN-... entry with Source: simplify-and-harden.

Promotion threshold: Recurrence-Count >= 3, seen in 2+ sessions, within 30-day window.

Meta-specific actions:

• Verbose prompt file → compress (target: 50% token reduction)
• Conflicting rules → consolidate to single source of truth
• Ambiguous instruction → rewrite with concrete examples or thresholds
• Hook producing wrong output → fix handler code and add test case

Periodic Review

Review .learnings/ at natural breakpoints:

When to Review

• Before publishing a new skill
• After modifying any core prompt file (AGENTS.md, SOUL.md, TOOLS.md, MEMORY.md)
• When adding or removing hooks
• When agent behavior degrades unexpectedly
• Monthly audit of all prompt files for staleness and conflicts

Quick Status Check

# Count pending meta issues
grep -h "Status\*\*: pending" .learnings/*.md | wc -l

# List pending high-priority infrastructure issues
grep -B5 "Priority\*\*: high" .learnings/META_ISSUES.md | grep "^## \["

# Find learnings for a specific area
grep -l "Area\*\*: hook_scripts" .learnings/*.md

# Find all rule conflicts
grep -B2 "rule_conflict" .learnings/LEARNINGS.md | grep "^## \["

Review Actions

• Resolve fixed infrastructure issues
• Promote recurring patterns to prompt files
• Link related entries across files
• Extract skill candidates (3+ related learnings)
• Prune stale memory entries referenced in learnings

Detection Triggers

Automatically log when you encounter:

Instruction Problems (→ learning with instruction_ambiguity):

• Agent says "I'm not sure what this instruction means" or paraphrases incorrectly
• Agent ignores a rule that should apply to the current task
• Agent asks for clarification on the same topic across multiple sessions
• Skill frontmatter description doesn't match actual behavior
• Rule file has no clear trigger condition

Configuration Conflicts (→ learning with rule_conflict):

• Two prompt files give contradictory guidance on the same topic
• CLAUDE.md convention conflicts with AGENTS.md workflow step
• .cursor/rules/ entry contradicts AGENTS.md instruction
• Skill description conflicts with global behavioral rules

Context Issues (→ learning with context_bloat):

• Agent loads too much context and truncates important information
• Prompt file exceeds 200 lines without structured formatting
• Same information duplicated across 2+ prompt files
• Historical context that belongs in README or changelog

Memory Problems (→ learning with prompt_drift):

• Memory entry references deleted files or outdated conventions
• Daily memory file contains information superseded by newer entries
• MEMORY.md rotation policy hasn't been applied in 30+ days
• Agent behavior changed because stale memory overrides current instructions

Hook Failures (→ meta issue):

• Hook output is empty when it shouldn't be
• Hook script exits with non-zero code
• Hook fires on wrong event or doesn't fire at expected event
• Hook configuration path points to non-existent script

Skill Problems (→ meta issue):

• Skill doesn't activate despite matching trigger conditions
• Skill frontmatter is malformed (missing name, description, or YAML errors)
• Skill template is missing required sections
• Skill extraction script produces invalid scaffold

Priority Guidelines

Area Tags

Use to filter learnings by infrastructure domain:

Best Practices

1. Keep prompt files concise — every token counts against context window
2. One authoritative source per rule — no duplication across files
3. Test skills in fresh sessions before publishing
4. Version prompt files — track what changed and when
5. Audit hooks monthly — silent failures are the worst failures
6. Prune memory aggressively — stale entries cause drift
7. Write skill descriptions as trigger conditions, not feature lists
8. Use structured formats (tables, lists) over prose in prompt files
9. Quote the problematic instruction verbatim when logging ambiguity
10. Log first, patch only on request — never silently change prompt files, hooks, memory, or skills

Hook Integration

Enable reminders through agent hooks only when needed. This is opt-in.

Hooks persist across sessions once installed. Keep them project-scoped. Do not install user-level or global hooks. Never use an empty matcher. PostToolUse inspects command output in-process; do not log raw output, secrets, or transcripts.

Conservative Mode (Recommended)

• Default to no hooks and log manually.
• If reminders are useful, enable UserPromptSubmit with scripts/activator.sh only.
• Enable PostToolUse (scripts/error-detector.sh) only in trusted environments when you explicitly want command-output pattern checks.

Quick Setup (Claude Code / Codex)

Create .claude/settings.json in your project:

{
  "hooks": {
    "UserPromptSubmit": [{
      "matcher": "self-improving-meta|prompt-governance|stackability|instruction.ambiguity|rule.conflict|context.bloat",
      "hooks": [{
        "type": "command",
        "command": "./skills/self-improving-meta/scripts/activator.sh"
      }]
    }]
  }
}

This injects an infrastructure-focused learning evaluation reminder after matching prompts (~50-100 tokens overhead).

Advanced Setup (With Error Detection)

{
  "hooks": {
    "UserPromptSubmit": [{
      "matcher": "self-improving-meta|prompt-governance|stackability|instruction.ambiguity|rule.conflict|context.bloat",
      "hooks": [{
        "type": "command",
        "command": "./skills/self-improving-meta/scripts/activator.sh"
      }]
    }],
    "PostToolUse": [{
      "matcher": "Bash",
      "hooks": [{
        "type": "command",
        "command": "./skills/self-improving-meta/scripts/error-detector.sh"
      }]
    }]
  }
}

Enable PostToolUse only if you want the hook to inspect command output for infrastructure-related error terms (hook, skill, AGENTS.md, frontmatter, truncated, etc.).

Available Hook Scripts

See references/hooks-setup.md for detailed configuration and troubleshooting.

Automatic Skill Extraction

Extracted skills are untrusted until a human reviews the generated SKILL.md. Do not keep or publish an extracted skill without explicit user approval.

When a meta-learning is valuable enough to become a reusable skill, extract it. Meta-skills extracting from meta-learnings is recursive improvement — the system improving itself.

Skill Extraction Criteria

Extraction Workflow

1. Identify candidate: Learning meets extraction criteria
2. Preview, then write only if the user approved --write in this session:

   ./skills/self-improving-meta/scripts/extract-skill.sh skill-name
   ./skills/self-improving-meta/scripts/extract-skill.sh skill-name --write
   

3. Customize SKILL.md: Fill in template with infrastructure-specific content
4. Update learning: Set status to promoted_to_skill, add Skill-Path
5. Test in fresh session: Meta-skills require fresh session verification because they modify shared infrastructure
6. Verify no regressions: Ensure other skills and hooks still work

Extraction Detection Triggers

In conversation: "This keeps happening", "Save this infrastructure fix as a skill", "Every project has this prompt file issue", "This hook pattern should be standard".

In entries: Multiple See Also links, high priority + resolved, rule_conflict or context_bloat with broad applicability, same Pattern-Key across projects.

Meta-Improvement Loop

This skill logs infrastructure problems. It does not apply edits by default:

1. Observe: Agent misinterprets instruction / hook fails / context is bloated
2. Log: Record the infrastructure issue in .learnings/
3. Analyze: Identify root cause (ambiguity, conflict, bloat, drift, gap, failure)
4. Propose: Draft a minimal patch. Do not apply it unless the user asked in this session.
5. Verify: If applied after approval, confirm the fix in a fresh session
6. Learn from the fix: If the fix itself causes issues, log that too

Do not treat logging as authorization to change shared prompt files, hooks, memory, or skills.

Multi-Agent Support

Gitignore Options

Keep learnings local (per-developer):

.learnings/

Track learnings in repo (team-wide): Only track .learnings/ after a human has reviewed entries for secrets, PII, and privileged content.

Hybrid (track templates, ignore entries):

.learnings/*.md
!.learnings/.gitkeep

Stackability Contract (Standalone + Multi-Skill)

This skill is standalone-compatible and stackable with other self-improving skills.

Namespaced Logging (recommended for 2+ skills)

• Namespace for this skill: .learnings/meta/
• Keep current standalone behavior if you prefer flat files.
• Optional shared index for all skills: .learnings/INDEX.md

Required Metadata

Every new entry must include:

**Skill**: meta

Hook Arbitration (when 2+ skills are enabled)

• Use one dispatcher hook as the single entrypoint.
• Dispatcher responsibilities: route by matcher, dedupe repeated events, and rate-limit reminders.
• Suggested defaults: dedupe key = event + matcher + file + 5m_window; max 1 reminder per skill every 5 minutes.

Narrow Matcher Scope (meta)

Only trigger this skill automatically for explicit meta signals such as:

• self-improving-meta|prompt-governance|stackability
• instruction.ambiguity|rule.conflict|context.bloat
• explicit meta intent in the user prompt

Do not match generic words such as hook, skill, rule, or prompt by themselves.

Cross-Skill Precedence

When guidance conflicts, apply:

1. security
2. engineering
3. coding
4. ai
5. user-explicit domain skill
6. meta as tie-breaker

Ownership Rules

• This skill writes only to .learnings/meta/ in stackable mode.
• It may read other skill folders for cross-linking, but should not rewrite their entries.
• Standalone mode writes to this project's .learnings/*.md log files only.
• Stackable mode writes only to the namespaced folder above and must not rewrite other skills' log entries.
• Promotion into AGENTS.md, SOUL.md, TOOLS.md, MEMORY.md, rules, hooks, or generated skills is not a logging write. Show a reviewed diff and apply only after explicit user approval.