skill-creator-thepexcel（技能创建器 thepexcel）

Skill Creator (ThepExcel Edition)

Based on Anthropic's official skill-creator (Apache 2.0). Enhanced with ThepExcel deployment workflow and structured audit/enhancement pipeline.

Quick Start

Figure out where the user is in this process and jump in. If they already have a draft, go straight to eval/iterate. If they say "just vibe with me", do that.

Core Principles

Concise is key — Context window เป็นทรัพยากรที่แชร์กัน ทุกบรรทัดต้องจ่ายค่า token

Explain the why — อธิบาย *ทำไม* ไม่ใช่แค่ *อะไร* Claude ฉลาดพอ generalize จาก reasoning ดีกว่า rule แข็งทื่อ ถ้าจะเขียน ALWAYS/NEVER all caps → yellow flag: reframe เป็น reasoning แทน

Generalize, don't overfit — Test cases คือตัวอย่าง ไม่ใช่ spec ทั้งหมด หา pattern ที่ work broadly

Keep the prompt lean — อ่าน transcript จริง ถ้า skill ทำให้ waste time → ตัดออก

Degrees of Freedom:

Skill Structure

skill-name/
├── SKILL.md (required)     ← < 500 lines ideal
│   ├── YAML frontmatter    ← name + description (+ compatibility optional)
│   └── Markdown body
├── agents/                 ← subagent instructions (grader, comparator, analyzer)
├── eval-viewer/            ← generate_review.py + viewer.html
├── assets/                 ← eval_review.html template
├── scripts/                ← deterministic code (execute without loading into context)
├── references/             ← loaded on demand (one level deep only)
└── evals/                  ← evals.json + test files

Progressive Disclosure (3 levels):

1. Metadata (name + description) — always in context (~100 words)
2. SKILL.md body — when triggered (< 500 lines)
3. Bundled resources — as needed

Frontmatter: name max 64 chars kebab-case, description max 1024 chars third person what + when

"When to use" in body = useless — Claude sees description only when deciding to trigger. Put all trigger context there.

Description tip: Make it slightly "pushy" to combat undertriggering — include specific contexts even if not explicitly named.

For anti-patterns: anti-patterns.md

Creating a Skill

Step 1: Capture Intent

ถามถ้าไม่ชัด หรือ extract จาก conversation ถ้ามีอยู่แล้ว:

1. Skill ควรทำอะไร?
2. Trigger เมื่อไหร่? (user phrases/contexts)
3. Expected output format?
4. ต้องการ test cases ไหม? — Skills ที่มี objective output ควรมี (file transforms, data extraction, fixed workflows) / Subjective skills (writing style, art) ไม่จำเป็น

Step 2: Interview & Research

ถาม edge cases, input/output formats, example files, dependencies Check available MCPs — research via subagents ถ้าทำได้ ใช้ /extract-expertise สำหรับ domain ที่ซับซ้อน

Step 3: Initialize

scripts/init_skill.py <skill-name> --path <output-directory>

Step 4: Write SKILL.md

ใส่: name, description (primary trigger mechanism — all "when to use" goes here), instructions

Writing style: imperative form, explain *why* behind each instruction, not rigid rules. Use theory of mind.

Design references:

• workflows.md — Sequential, conditional, loops
• output-patterns.md — Templates, formatting
• anti-patterns.md — Common mistakes

Step 5: Write Test Cases

2-3 realistic prompts — the kind of thing a real user would actually say. Share with user for confirmation. Save to evals/evals.json (don't add assertions yet):

{
  "skill_name": "example-skill",
  "evals": [
    {
      "id": 1,
      "prompt": "User's task prompt",
      "expected_output": "Description of expected result",
      "files": [],
      "expectations": []
    }
  ]
}

See references/schemas.md for full schema including assertions field.

Running and Evaluating Test Cases

This is one continuous sequence — don't stop partway. Do NOT use any other testing skill.

Put results in <skill-name>-workspace/ (sibling to skill directory), organized by iteration-N/eval-name/.

Step 1: Spawn All Runs in the Same Turn

For each test case, launch two subagents simultaneously — one with-skill, one baseline. Don't launch with-skill first and come back for baselines later.

• New skill: baseline = no skill at all, save to without_skill/outputs/
• Improving existing skill: baseline = old version snapshot (cp -r <skill-path> <workspace>/skill-snapshot/), save to old_skill/outputs/

Write eval_metadata.json for each eval:

{
  "eval_id": 0,
  "eval_name": "descriptive-name",
  "prompt": "The task prompt",
  "assertions": []
}

Step 2: While Runs Are in Progress, Draft Assertions

Don't wait idle. Draft objectively verifiable assertions with descriptive names (they appear in the viewer). Explain them to the user. Subjective skills → qualitative only, skip assertions.

Update eval_metadata.json and evals/evals.json with assertions.

Step 3: Capture Timing Data

When each subagent completes, save immediately — this data exists only in the task notification:

{"total_tokens": 84852, "duration_ms": 23332, "total_duration_seconds": 23.3}

Save to timing.json in the run directory.

Step 4: Grade → Aggregate → Analyze → Launch Viewer

1. Grade — spawn grader subagent using agents/grader.md, save to grading.json. For assertions checkable programmatically, write a script rather than eyeballing.
2. Aggregate — run from skill-creator directory: python -m scripts.aggregate_benchmark <workspace>/iteration-N --skill-name <name> Produces benchmark.json and benchmark.md with pass_rate, time, tokens (mean ± stddev + delta).
3. Analyst pass — read agents/analyzer.md (Analyzing Benchmark Results section) to surface patterns aggregate stats hide: non-discriminating assertions, high-variance evals, time/token tradeoffs.
4. Launch viewer: nohup python <skill-creator-path>/eval-viewer/generate_review.py \ <workspace>/iteration-N \ --skill-name "my-skill" \ --benchmark <workspace>/iteration-N/benchmark.json \ > /dev/null 2>&1 & VIEWER_PID=$! Iteration 2+: add --previous-workspace <workspace>/iteration-<N-1> Cowork/headless: use --static <output_path> instead. Feedback downloads as feedback.json. ⚠️ GENERATE THE EVAL VIEWER BEFORE evaluating inputs yourself. Get results in front of the human first.
5. Tell the user: "ผลอยู่ในเบราว์เซอร์แล้วค่ะ — tab Outputs ดู output แต่ละ test case, tab Benchmark ดู metrics เมื่อดูเสร็จแล้วกลับมาบอกหนูได้เลย"

Step 5: Read Feedback

Read feedback.json. Empty feedback = satisfied. Focus on test cases with specific complaints.

Kill viewer: kill $VIEWER_PID 2>/dev/null

Improving the Skill

How to Think About Improvements

1. Generalize — หา pattern ที่ work broadly ไม่ใช่ fix specific test case
2. Keep lean — อ่าน transcript จริง ถ้า skill ทำให้ waste time → ตัดออก
3. Explain the why — ถ้าจะเขียน ALWAYS/NEVER → yellow flag: reframe เป็น reasoning แทน
4. Bundle repeated work — ถ้า 3 test cases ล้วน write create_docx.py เอง → bundle ใน scripts/

Iteration Loop

1. Apply improvements to skill
2. Rerun all test cases into iteration-<N+1>/ (including baselines)
3. Launch viewer with --previous-workspace pointing at previous iteration
4. Wait for user review → read feedback → repeat

Stop when: user satisfied / all feedback empty / not making meaningful progress.

For quick fixes (typo, small gap): direct edit → skip eval loop.

Blind Comparison (Advanced)

For rigorous version comparison, read agents/comparator.md + agents/analyzer.md for blind A/B judgment. Optional — the human review loop is usually sufficient.

Enhancement Mode

ใช้เมื่อต้องการปรับปรุง skill เดิมที่มีอยู่แล้ว แบบ structured — เหมาะเมื่อยังไม่รู้ว่า skill มีปัญหาตรงไหน หรืออยากยก quality ขึ้นเป็น systematic

Route

Full Pipeline: AUDIT → RESEARCH → INTEGRATE → OPTIMIZE → VALIDATE

AUDIT

อ่าน target skill ทั้งหมด → score ด้วย audit-rubric.md:

SKILL: [name]
SCORES: Coverage [?] | Depth [?] | Structure [?] | Actionability [?] | Examples [?]
TOTAL: [?]/25 → [Draft/Working/Solid/Production]

จุดที่ควรปรับ:
1. [ปัญหา + ผลกระทบ]

→ ถามผู้ใช้ก่อน: "ปรับทั้งหมด หรือเลือกเฉพาะข้อ?"

RESEARCH

ใช้ /deep-research หรือ /extract-expertise เพื่อเติม knowledge gaps

INTEGRATE

Classify findings → prioritize by impact → merge ด้วย integration-patterns.md

OPTIMIZE

Apply skill-creator standards: progressive disclosure, conciseness, references/

VALIDATE

Before/after comparison:

| Dimension | Before | After | เปลี่ยนอะไร |
|-----------|--------|-------|------------|

Log ใน enhancement-log.md → จากนั้นรัน Eval loop เพื่อยืนยัน improvement จริง

Benchmark & Description Optimization

Benchmark

Rerun all evals 3x per configuration with aggregate_benchmark.py. Track pass rate, time, tokens across iterations/models. ใช้สำหรับ regression detection เมื่อ model อัปเดต หรือหลัง enhance

Description Optimization

หลัง skill เสร็จ เสนอให้ optimize description สำหรับ triggering accuracy ที่ดีขึ้น

Step 1: Generate 20 trigger eval queries (mix should/should-not trigger). Be realistic — personal context, file paths, casual speech, typos. Near-miss negatives are the most valuable test cases.

Present via HTML template:

1. Read assets/eval_review.html, fill: __EVAL_DATA_PLACEHOLDER__, __SKILL_NAME_PLACEHOLDER__, __SKILL_DESCRIPTION_PLACEHOLDER__
2. Write to /tmp/eval_review_<skill-name>.html → open it
3. User edits, clicks "Export Eval Set" → ~/Downloads/eval_set.json

Step 2: Run optimization loop (background):

python -m scripts.run_loop \
  --eval-set <path-to-trigger-eval.json> \
  --skill-path <path-to-skill> \
  --model <model-id-powering-this-session> \
  --max-iterations 5 \
  --verbose

Splits 60/40 train/test, iterates up to 5x, returns best_description selected by test score to avoid overfitting.

Step 3: Apply best_description to SKILL.md frontmatter. Show before/after + scores.

Note: requires claude -p CLI → Claude Code only, not Claude.ai

Deploy (ThepExcel)

┌─ Skill ใหม่
│   ├─ Public?   → /mnt/d/agent-skills/[skill-name]/
│   └─ Private?  → /mnt/d/claude-private/skills/[skill-name]/
│
├─ Symlink
│   └─ ln -s /mnt/d/[repo]/[skill-name] ~/.claude/skills/[skill-name]
│
├─ Update registry
│   └─ /mnt/d/claude-master/CLAUDE.md → Skills Inventory
│
└─ Commit & Push
    └─ git add → commit → push (ทั้ง skill repo + claude-master)

Validate before deploy:

scripts/quick_validate.py <path/to/skill-folder>
scripts/package_skill.py <path/to/skill-folder>

Platform Notes

Claude.ai: skip baselines + benchmarking, run test cases yourself one at a time, show results + ask feedback inline.

References

Related Skills

• /extract-expertise — Extract expert knowledge to inform skill content
• /deep-research — Research domain before building or enhancing skill
• /optimize-prompt — Optimize skill descriptions and system prompts