Skill Curator: When Your Agent Has 125 Skills but Only 64 Survived

Premise: You installed 125 skills for your Agent. You think it is powerful. In reality, it can correctly invoke only 64 of them. Of the other 61, 6 can never be triggered, and 55 are basically ineffective for Chinese users.
Solution: Skill Curator, a six-stage automated curation engine. Scan → Diagnose → Adapt → Scenario → Report → Execute.
Result: 125 skills improved from 51% health to 99.2%. 6 critical → 1 (ghost skill), 55 warnings → 0, 64 healthy → 124.

🎭 Chapter 1: The Library Paradox

Imagine you walk into a library. The shelves are packed with books, 125 of them. But:

60% of the books have no Chinese catalog, so you cannot find them when searching in Chinese.
6 books do not even have covers, so you have no idea what is inside.
15 books overlap completely in content, but you do not know which one to discard.
No one tells you how to use each book; you stand at the entrance, at a loss.

This is the reality every AI Agent faces.

We call this phenomenon the "Library Paradox": More skills ≠ greater capability. Without curation, the number of skills is inversely proportional to usability.

The Root of the Problem

The Agent ecosystem has three structural problems:

Skills are downloaded, not adapted. In open-source skills, 95% of the description field is pure English. A Chinese user says, in Chinese, “help me build a website” → the description contains only "frontend design" and "UI engineering" → keyword matching fails → the skill is effectively useless.
Formatting errors are inherent. Community skills are copied from one another, with misaligned frontmatter YAML, missing required fields, and encoding issues. A blank description can prevent a skill from ever being triggered.
No one actively maintains skills. Users download skills once and never maintain them. You need an automated "skill curator."

🔬 Chapter 2: Audit, Blood Test Report for 125 Skills

In May 2026, we performed the first comprehensive health scan of UltraClaw's 125 skills. Here are the blood test results:

Health Distribution

🔴 Fatal

No description or corrupted format, permanently cannot be triggered. Equivalent to "this book has no cover".

🟡 Warning

Missing Chinese keywords, description is English only. Chinese users have a trigger success rate below 20%.

🟢 Healthy

Complete format + bilingual Chinese and English keywords. But there are still functional overlaps and untested risks.

Core Findings

Metric	Data	Severity
Overall health rate	51.2%	Half of the skills are subhealthy or dead
Non-English trigger success rate	~20%	Only 1 in 5 Chinese commands can find the right skill
description language coverage	1.2 languages/skill	95% of skills have only English description
Functionally overlapping skill groups	7 groups (15 skills)	The same function has 2-3 overlapping skills that interfere with each other

Conclusion: You think your Agent is strong. In reality, one out of every two skills is slacking off.

🧬 Chapter 3: Six-Phase Lifecycle

The core design philosophy of Skill Curator is full lifecycle management. It is not a one-time fix, but ongoing health maintenance.

{["1️⃣ Scan", "2️⃣ Diagnose", "3️⃣ Adapt", "4️⃣ Scenarios", "5️⃣ Report", "6️⃣ Execute"].map((phase, i) => (

{phase}

))}

Phase 1: Scan, Full Health Inventory

The scanner traverses all SKILL.md files under the skills/ directory and extracts:

Frontmatter completeness: Whether the required fields name and description are present
Language coverage: How many languages are included in description
Keyword matrix: Which trigger words exist for each language
File metadata: Size, modification time, and license type

# Scan report example snippet
{
  "skill": "frontend-design",
  "health": "🟡 warning",
  "issues": [
    "description contains only English (1/6 language)",
    "Missing keywords: website, frontend, website building, webpage"
  ],
  "recommended_action": "inject_keywords"
}

Scan results automatically generate a JSON report for consumption by downstream phases.

Phase 2: Diagnose, Three-Tier Classification

This is the most valuable design in Skill Curator: not all problems are equally important.

🔴 Critical

The skill cannot be triggered at all. It requires immediate repair.

The description field is missing or empty.
Frontmatter YAML syntax errors cause parsing to fail.
The SKILL.md file is corrupted or empty.

The first scan found 6 critical skills. Of these, 4 were due to abnormal frontmatter formatting in community skills (incorrect indentation, use of non-standard fields causing the parser to skip required fields).

🟡 Warning

The skill exists but is extremely inefficient. It is nearly unusable for non-English users.

description contains only English (trigger words cover 1/6 of languages).
Missing keywords in Traditional Chinese, Simplified Chinese, Japanese, Korean, and Arabic.
Frontmatter is complete, but description is too short (< 50 characters).

The first scan found 55 warning skills. This is the largest problem set, accounting for 44% of the total. Each is "theoretically usable," but for a boss who uses Traditional Chinese, the trigger rate is below 20%.

🟢 Suggestion

The skill functions normally but has room for optimization. It requires user confirmation before execution.

Functional overlap: multiple skills implement the same or similar functionality (such as 3 different "frontend design" skills).
Never used: the skill has been installed for more than 30 days but has zero invocation records.
Token optimization: the skill's SKILL.md is too large (> 20KB), consuming a large amount of context on every load.

The first scan found 7 groups of functional overlap (involving 15 skills), as well as 12 "ghost skills" (installed but never invoked).

Phase 3: Adapt, Automated Repair Engine

This is the "surgical operation" phase of Skill Curator. It runs fully automatically, but always backs up first.

3.1 Six-Language Keyword Injection

Core algorithm: parse the semantic content of description → generate corresponding keywords for each missing language → inject them into the description field.

Before fix:
description: "Full-stack web application development with modern frameworks"

After fix:
description: "Full-stack development Website creation Frontend and backend Web application development Full-stack development"
تطوير الويب Full-stack development site web Full-stack web application development 
with modern frameworks"

Six-language coverage:

🇹🇼 Traditional Chinese
🇨🇳 Simplified Chinese
🇯🇵 Japanese
🇰🇷 한국어
🇸🇦 العربية
🇬🇧 English (original)

3.2 Frontmatter Repair

Automatically fix common formatting corruption:

YAML indentation error → Reformat
Missing description → Automatically extract and generate from the skill content
Non-standard fields → keep but move to comment area
Encoding issues → standardize to UTF-8

3.3 Safety Mechanism: Always Back Up First

# Run automatically before each fix
cp skills/frontend-design/SKILL.md skills/frontend-design/SKILL.md.bak.$(date +%s)

Result: In the first curation pass, 61 skills were automatically repaired. Zero errors. Because the original files were preserved before every modification.

Phase 4: Scenario Generation Scenarios, "This Is How You Should Use Me"

Fixing only makes skills discoverable. Scenario generation enables skills to be used correctly.

Automatically generate 3-5 trigger scenarios for each skill:

## 🧪 Trigger Scenarios

| What you want to do | What you should say | Trigger Skill |
|-----------|---------|---------|
| Build a company website | "Help me make an official website" | Frontend Design |
| Stock research | "Research Tencent" | AK-HK-Stock-DD |
| Generate an investment proposal | "Write an investment proposal" | AK-Investment-Proposal |
| Send daily report email | "Send today's report" | Email Report |
| Create a Feishu document | "Create meeting minutes" | Feishu Create Doc |

The Value of Scenario Generation:

Lower learning costs: Users do not need to know skill names, only "what I want to do"
Improve trigger accuracy: Trigger words in the scenario are synced to description, creating a positive feedback loop
Discover hidden features: Users may not know that the Agent has certain skills, and scenario generation surfaces them

Phase 5: Report, Transparency Is the Cornerstone of Trust

After curation is complete, a structured report is automatically generated and pushed to Feishu:

## 📊 Skill Curator Curation Report: 2026-06-10

### Health Trend
- Overall health rate: 51.2% → 99.2% (+48.0%)
- Fatal skills: 6 → 1 (-83.3%)
- Warning skills: 55 → 0 (-100%)
- Healthy skills: 64 → 124 (+93.8%)

### Actions Performed This Run
- 🔧 Auto repair: 61 skills
- 🌏 Keyword injection: 61 skills (average +15 keywords each)
- 🔨 Frontmatter repair: 4 skills
- 📋 Scenario generation: 125 skills (487 scenarios total)
- ⚠️ Recommended uninstall: 3 overlapping skills (pending approval)
- 👻 Ghost skills: 1 (cannot be repaired, marked)

### Language Coverage Improvement
- Before repair: average 1.2 languages/skill
- After repair: average 5.8 languages/skill

Report Design Principles:

Not a technical log for developers, but a health report for decision-makers
Every operation is traceable, including backup file paths, modified content, and timestamps
Advisory operations require explicit user approval, with no black-box decisions

Phase 6: Execute, Decision Boundaries for Human-AI Collaboration

This is Skill Curator's most critical design principle: the boundaries of automation.

Operation Type	Execution Method	Reason
Keyword injection	🤖 Fully automatic	No risk, reversible, backed up
Frontmatter repair	🤖 Fully automatic	Format correction, no semantic changes
Scenario generation	🤖 Fully automatic	Additional content, does not modify original files
Deduplication merge	👤 Requires approval	Involves functional changes
Skill uninstallation	👤 Requires approval	Irreversible operation
Routing rule adjustment	👤 Requires approval	Affects global behavior

Design philosophy: Automate what can be automated; clearly mark what cannot be automated and wait for the user to decide. An Agent should not make value judgments on behalf of humans.