Build Your Own Agent Skill Interactively with Anthropic's skill-creator

## What is it, and what problem does it solve? If you find yourself re-explaining the same workflow, the same set of rules, and the same steps to Claude at the start of every new conversation, what you need is an "Agent Skill." A skill is a folder containing instructions (SKILL.md), scripts, and resources that Claude "dynamically loads" when it detects a relevant context, sparing you from repeating yourself every time. The catch is that hand-crafting a skill that's both useful and "triggers precisely" isn't easy — a description that's too narrow means it never fires, and one that's too broad means it fires everywhere. skill-creator is Anthropic's official "meta-skill" (a skill for making skills) built to solve exactly this problem. ## Why this source is worth using This is a first-party implementation from Anthropic's official `anthropics/skills` repo, not third-party guesswork. It lays bare Anthropic's internal standard process for building skills: a fixed directory structure, YAML frontmatter conventions, a three-tier progressive-disclosure loading model, and a methodology that actually runs A/B evals. The official frontmatter describes it as: "Create new skills, modify and improve existing skills, and measure skill performance." The repo ships real scripts including `scripts/run_eval.py`, `scripts/aggregate_benchmark`, `scripts/run_loop`, `scripts/package_skill`, and `eval-viewer/generate_review.py`, plus grading subagents such as `agents/grader.md`, `agents/comparator.md`, and `agents/analyzer.md`. ## How to use it (key points) 1. State the skill you want to build in one sentence: what it does, when it should trigger, and what the expected output looks like. 2. Let it interview you to fill in edge cases and input/output formats. 3. It produces a `SKILL.md` (name + description required) plus optional scripts/references/assets directories. 4. The key principle is "progressive disclosure": keep the SKILL.md body under roughly 500 lines, push detailed documentation into references/ and long scripts into scripts/, so Claude only loads them on demand and saves context. 5. Validate with evals: in the same run, dispatch two subagents — "with the skill" versus "without it" — side by side, then iterate on the description's trigger accuracy. ## When to use it - You have a highly repetitive, clearly-ruled workflow you want to turn into a standing capability for Claude. - You already have a skill, but it keeps "not triggering when it should" or "triggering randomly," and you want to quantify and improve its description with evals. - You want to package your team's SOPs or standards into a shareable `.skill` file. ## Note The description needs to "actively list concrete trigger scenarios" rather than just describe functionality — that's what determines whether it activates correctly. The official docs also caution that the repo's content is "for demonstration and educational purposes only," and actual Claude behavior may differ slightly from what's shown. Source: anthropics/skills (by Anthropic, Apache-2.0 license) — this entry is an edited adaptation with rewritten content; see the link above for the original.

填下面的欄位，上方 prompt 會即時替換 [方括號] 內容。填好後按「複製組好的 prompt」直接丟進工具。

你是一位 Agent Skill 設計顧問。請依下列流程，互動式地幫我打造一個全新的 Claude Skill。

# 目標技能
- 技能要做的事：{{SKILL_PURPOSE}}
- 觸發情境（使用者說什麼 / 在做什麼時該啟動）：{{TRIGGER_CONTEXT}}
- 期望輸出：{{EXPECTED_OUTPUT}}

# 流程（逐步進行，每步先問清楚再動手）
1. 捕捉意圖：用你的話複述這個技能的「完成定義」，確認後再繼續。追問邊界情況、輸入/輸出格式、成功標準。
2. 產出目錄結構：
   skill-name/
   ├── SKILL.md            # 必填：YAML frontmatter + Markdown 指示
   ├── scripts/            # 選填：可執行程式
   ├── references/         # 選填：詳細文件（按需載入）
   └── assets/             # 選填：範本 / 圖示 / 字型
3. 寫 SKILL.md 的 YAML frontmatter，只需 name 與 description 兩個必填欄位：
   - name: 技能識別名（kebab-case）
   - description: 「做什麼 + 何時用」一起寫，要『主動 pushy』地列出具體觸發情境，避免技能被低估而不啟動。
4. 寫 SKILL.md 本文：遵守漸進式揭露三層——metadata 永遠載入、SKILL.md 本文觸發時載入（目標 < 500 行）、references/scripts 按需載入。把長文件、長範例移到 references/。
5. 設計評測：在 evals/evals.json 列出測試案例（id / prompt / expected_output / assertions）。
6. 評測方法：同一回合派出兩個 subagent——一個有此技能、一個沒有——並排比較輸出，先給人看再自動評分。
7. 依回饋迭代：從具體抱怨歸納通則，不要做過度擬合的瑣碎修補；換比喻、換結構，保持技能在多種真實情境下都能轉移。
8. 優化 description：針對「該觸發卻沒觸發 / 不該觸發卻觸發」反覆調整，直到觸發精準。
9. 打包：整理成最終 .skill 檔。

請先做第 1 步：複述完成定義並列出你要追問我的問題。