🌿 Skill Garden — Skill Evolution Engine

"Every skill should get better the more you use it."

Philosophy

Skill Garden treats skill improvement as a continuous, invisible process — not a special operation. It runs passively in the background, accumulating observations from every skill invocation, then periodically synthesizes them into concrete improvements.

Key design principles:

• Token-efficient: Lightweight structured logs, batch processing, no real-time overhead
• User-in-control: High-confidence changes auto-apply; uncertain ones ask first
• Transparent: Every change is explained, nothing happens silently without explanation
• Self-contained: Manages its own memory, dashboard, proposals, and cron schedule

Three-Layer Architecture

Layer 1 — Passive Observation (near-zero token cost)
  Every skill invocation → 1-line structured log entry
  Abnormal outcomes (FAIL/SLOW) → detailed log with evidence

Layer 2 — Weekly Batch Analysis (isolated agent, ~5-15 min)
  Read all accumulated logs
  Run evaluation engine across 6 dimensions
  Generate specific improvement proposals

Layer 3 — Targeted Modification (low frequency, high precision)
  Confidence ≥ 90% → apply immediately, notify user
  Confidence 70–89% → apply with [experimental] tag
  Confidence 50–69% → write to proposals, ask user
  Confidence < 50% → log as observation only

When This Skill Activates

Trigger 1 — After every skill use (automatic, passive): When any skill finishes executing, immediately log the outcome using the format in references/usage_tracker.md. This is the most important layer — it costs almost nothing and feeds everything else.

Trigger 2 — On user request:

• "grow this skill" / "improve skill" / "optimize skill"
• "why did this fail" / "analyze this skill"
• "run skill analysis" / "check my skills"
• "skill health" / "skill dashboard"

Trigger 3 — On schedule (automatic, every Sunday 20:00): An isolated agent runs batch_analyze.py and generate_report.py, applies high-confidence improvements, and sends you a summary.

Layer 1: Passive Observation

After any skill finishes (any outcome: OK, FAIL, PARTIAL, SLOW, SKIP), immediately write a structured log. Use log_insight.py or write directly to the skill's references/usage_log.md.

Log Entry Format

For OK outcomes with nothing notable (minimal tokens):

## YYYY-MM-DD HH:MM
Trigger: [trigger in ≤10 words]
Outcome: OK
Signal: [one-line finding or "No issues"]

For PARTIAL, FAIL, SLOW outcomes (always log all fields):

## YYYY-MM-DD HH:MM

### Trigger
[What the user asked for, ≤10 words]

### Outcome
OK | PARTIAL | FAIL | SLOW | SKIP

### Signal
[One specific phrase: what this tells us about the skill]
Examples:
  - "Covered: standard use case works perfectly"
  - "Missing: error handling for network timeouts"
  - "Ambiguous: step 3 could be interpreted two ways"
  - "Outdated: API version in skill doesn't match current"

### Evidence
[1-2 sentences. Quote or paraphrase exact output/error. Be specific.]

### Flags
[Comma-separated tags: [new_trigger] [missing_coverage] [confusing_step]
 [outdated_info] [token_heavy] [edge_case] [user_workaround_used]
 [config_stale] [api_change] [Covered] [success_boost]]

Using log_insight.py

# Quick OK log (minimal)
python3 ~/.openclaw/workspace/skills/skill-garden/scripts/log_insight.py \
  --skill github-trending-summary \
  --trigger "daily top 5 repos" \
  --outcome OK \
  --signal "Covered: standard case"

# Detailed failure log
python3 ~/.openclaw/workspace/skills/skill-garden/scripts/log_insight.py \
  --skill banxuebang-helper \
  --trigger "check homework" \
  --outcome FAIL \
  --signal "Missing: semester selector not dynamic" \
  --evidence "Config hardcoded to 2024-2025 but API shows 2025-2026 is current." \
  --flags "missing_coverage,config_stale" \
  --mark-landmark "SkillImproved"

Rule of thumb: If you had to pause, reconsider, or work around something — log it with full detail. If it just worked perfectly — log minimally. The goal is signal, not noise.

Layer 2: Weekly Batch Analysis

Run manually or wait for the Sunday cron trigger.

Manual Trigger

Say: "run skill analysis" or "grow all skills"

The analysis does the following in order:

1. Scan all skills — read every references/usage_log.md
2. Evaluate each skill across 6 dimensions (see references/evaluation_engine.md)
3. Generate proposals — for each skill with score below threshold
4. Apply high-confidence changes — auto-edit SKILL.md for confident improvements
5. Update dashboard — rewrite references/dashboard.md
6. Notify user — send summary message

Running Scripts Directly

# Full batch analysis (evaluate all skills, generate proposals)
python3 ~/.openclaw/workspace/skills/skill-garden/scripts/batch_analyze.py

# Analyze one skill only
python3 ~/.openclaw/workspace/skills/skill-garden/scripts/batch_analyze.py --skill github-trending-summary

# Dry run (proposals only, don't apply)
python3 ~/.openclaw/workspace/skills/skill-garden/scripts/batch_analyze.py --dry-run --min-confidence 70

# Generate/refresh dashboard
python3 ~/.openclaw/workspace/skills/skill-garden/scripts/generate_report.py

# Output as JSON (for integrations)
python3 ~/.openclaw/workspace/skills/skill-garden/scripts/generate_report.py --output json

Layer 3: Applying Improvements

The Six Evaluation Dimensions

See references/evaluation_engine.md for the full evaluation algorithm, scoring thresholds, and confidence calibration guide.

Applying an Edit to SKILL.md

When a proposal meets the confidence threshold:

1. Read the current SKILL.md
2. Identify the exact text to replace using edit tool
3. Write the improved version
4. Add a brief changelog note at the top of the edit:

   <!-- Auto-improved by Skill Garden: YYYY-MM-DD
        Reason: [confidence]% confidence — [evidence summary] -->
   

5. Update references/improvement_proposals.md to mark as applied
6. Notify the user with a summary of what changed

Editing Checklist

Before applying any edit:

• Change is specific and testable (not vague advice)
• New text is more concrete than old text (examples > statements)
• If adding a step, verify it doesn't contradict existing steps
• If removing text, verify no other part of the skill depends on it
• If changing description, verify all log triggers are now covered
• Change addresses the flagged evidence, not just the symptom

Dashboard

The dashboard (references/dashboard.md) shows:

• Overall skill ecosystem health
• Per-skill scores across all 6 dimensions
• Recent signals and flags
• Pending proposals
• Weekly outcome distribution
• Recent landmark events

Regenerate with:

python3 ~/.openclaw/workspace/skills/skill-garden/scripts/generate_report.py

Cron Setup

Set up a weekly Sunday 20:00 analysis run:

openclaw cron add \
  --name "Skill Garden Weekly Analysis" \
  --schedule '{"kind":"cron","expr":"0 20 * * 0","tz":"Asia/Shanghai"}' \
  --sessionTarget isolated \
  --payload '{"kind":"agentTurn","message":"Run full skill analysis: execute batch_analyze.py then generate_report.py and notify me of any improvements applied.","timeoutSeconds":900}'

Or manually trigger with:

grow all skills

Reference Files

Examples in Context

Example 1: Passive Log — Skill Works Perfectly

User: "给我推荐一首今天听的歌" Skill: daily-song-recommender — succeeds, email sent Your action: Write one-line abbreviated log:

## 2026-04-22 09:15
Trigger: daily song recommendation
Outcome: OK
Signal: Covered: standard use case, no issues

Example 2: Passive Log — Skill Fails Due to Missing Coverage

User: "查一下这周的数学作业" Skill: banxuebang-helper — API error: semester field stale Your action: Write full structured log:

## 2026-04-22 21:30

### Trigger
banxuebang: check math homework

### Outcome
FAIL

### Signal
Missing: semester config hardcoded, not dynamic

### Evidence
API returned {"code": 500, "msg": "semester not found"}.
Config in skill has "2024-2025下学期" but actual is "2025-2026下学期".

### Flags
[missing_coverage] [config_stale]

Example 3: Passive Log — New Trigger Not in Description

User: "爬一下GitHub今日热榜发到我邮箱" Skill: github-trending-summary — works, but user's phrasing not in description Your action: Write log with [new_trigger] flag:

## 2026-04-22 10:00

### Trigger
github-trending-summary: crawl and email

### Outcome
OK

### Signal
Covered: but user said "爬" not in description triggers

### Evidence
Skill handled it fine, but description doesn't mention "爬" as a trigger phrase.

### Flags
[new_trigger]

Example 4: User Requests Analysis

User: "run skill analysis" Your action:

1. Run batch_analyze.py --dry-run
2. Read the proposals from output
3. Apply high-confidence changes (≥90%) via edit tool
4. Run generate_report.py to refresh dashboard
5. Message user: "Found N improvement(s) — applied X automatically, Y need your review"

Example 5: Weekly Cron Fires (Sunday 20:00)

Isolated agent runs full cycle:

1. batch_analyze.py scans all 20 installed skills
2. Finds github-trending-summary: 1 [new_trigger] flag, coverage 66%
3. Generates proposal with 65% confidence → written to proposals
4. Updates dashboard
5. You receive: "🌿 Weekly analysis done. github-trending-summary needs description update (65% confidence — needs more data to auto-apply). Review?"