Install any skill in seconds. Free to start, no credit card required.
Get Started Free →建立與更新符合專案最新規範(如 description 換行與觸發詞語法、title 欄位使用時機、標籤正規化及 references 目錄規劃)的高品質技能之微調版指南。 當使用者提及以下關鍵字或情境時觸發: - "建立新技能" 或 "建立技能" - "微調技能" 或 "微調技能指南" - "新增技能 / create skill" - "技能 Frontmatter 規範 / skill frontmatter" - "技能觸發詞 / skill triggers"
.claude/skills/bluelovers-skill-creator-fine-tuned/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 42% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 77% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 87% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 98% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 60% | 0% |
此技能是基於 skills/skill-creator 的微調版本,旨在引導 AI Agent 建立與維護高品質、標準化且符合專案最新 Frontmatter 與標籤規範的 Skill。
skill-creator 的差異本微調版技能在原版基礎上,強化了以下關鍵規範與實踐:
references/ 目錄:推動「漸進式揭露原則」,將複雜的語法細節、規範、範例與 schemas 抽離至 references/ 中,保持 SKILL.md 精簡。description 格式優化:規定在需要分行(如列表)時使用 |-,在不需要分行但原始碼需換行排版時使用 >-。obsidian-fm-tags,要求所有 .md 檔案(包含 main 檔案與 reference 檔案)都必須加上標準化 tags。title 欄位的使用時機:只有在檔名、第一個 Heading 與實際上表達的標題不一致時才加上 title 欄位。 skill-name/ ├── SKILL.md (必要,核心工作流程與導覽) └── references/ (必要,存放詳細規範、範例與 schemas) ├── frontmatter-specification.md └── other-references.md
本微調版的核心規範,詳細實作細節請查閱 frontmatter-specification.md。
|- 與 >-)在 SKILL.md 的 YAML Frontmatter 中,description 是最關鍵的觸發欄位,必須清楚且完整地包含觸發詞(Triggers)列表。為了使 Agent 精確識別觸發條件,應善用 YAML 的區塊文字折疊與保留語法:
|-(保留換行符,並刪除區塊末尾的換行符)。這是添加觸發詞列表時的標準格式。>-(將折疊的換行符換成空格,原始碼可自由換行,美化排版)。> !IMPORTANT] > 每個技能的主 SKILL.md 的 description 必須包含觸發詞列表(Triggers when user mentions...),以便引導 Agent 在對話中精確觸發該技能。因此,主技能的 description 應統一使用 |- 語法。
title 欄位當以下三者不一致時,在 frontmatter 中使用 title 欄位記錄實際標題:
hook-experimental-compaction-autocontinue.md)# Hook: experimental.compaction.autocontinue)yaml--- title: 深入分析 experimental.compaction.autocontinue 機制 tags: - opencode/plugin-system - opencode/hooks/autocontinue ---
title。title 僅用於需要不同於檔名與 heading 的自然語言表述時。參考 obsidian-fm-tags,Skill 目錄下的所有 .md 檔案都必須在 frontmatter 中包含標準化的 tags 欄位。
nodejs 而非 Node.js)。-) 分隔多詞(如 skill-creation)。/) 建立層級關係的巢狀標籤(如 agents/skills)。cpp 代替 C++)。references/ 資料夾references/ 目錄。references/ 中,保持 SKILL.md 簡潔。SKILL.md當編輯技能檔案時,請記住這是寫給「另一個 Agent」看的。撰寫時應遵循以下細部規則:
Frontmatter 決定了技能的觸發與標籤化,必須嚴格配置:
name:技能名稱。極為重要:YAML 中的技能名稱(name)必須與該技能的資料夾名稱完全一致(例如:資料夾名稱為 skill-creator-fine-tuned,則 Frontmatter 中 name 欄位的值必須完全為 skill-creator-fine-tuned,不可有大小寫、底線或連字號的任何差異,否則驗證與打包會失敗)。description:這是主要的觸發機制。SKILL.md 的 description 欄位一律必須使用 |- 語法以保留排版。tags:依據標籤正規化規則,為技能添加適當的標準化標籤。[frontmatter-specification.md](./references/frontmatter-specification.md))明確指向 references/ 中的參照檔案,指引 Agent 按需加載。references/ 中的每個 .md 檔案皆包含正確的 tags。title 欄位加以記錄。description 應使用 >- 語法進行美化排版。package_skill.py)驗證 YAML 格式、目錄結構與 description 的完整性,並打包為 .skill 檔分發。SKILL.md 中的觸發詞列表、工作流程指引或 references/ 中的範例,以持續提升 Agent 執行的確定性。| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 19,555 | 14,563 | -26% | 1 | 1 | 0% | 3,748 | 5,339 | +42% | 0 | 0 | — |
case-02 | fail→pass | 15,578 | 14,453 | -7% | 1 | 1 | 0% | 3,037 | 5,369 | +77% | 0 | 0 | — |
case-03 | fail→pass | 20,012 | 16,039 | -20% | 1 | 1 | 0% | 3,034 | 5,668 | +87% | 0 | 0 | — |
case-04 | pass→pass | 6,583 | 4,648 | -29% | 1 | 1 | 0% | 1,037 | 3,357 | +224% | 0 | 0 | — |
case-05 | fail→pass | 10,891 | 6,755 | -38% | 1 | 1 | 0% | 1,880 | 3,714 | +98% | 0 | 0 | — |
case-06 | pass→pass | 5,751 | 7,687 | +34% | 1 | 1 | 0% | 1,007 | 3,128 | +211% | 0 | 0 | — |
case-07 | pass→pass | 7,563 | 5,044 | -33% | 1 | 1 | 0% | 1,518 | 3,430 | +126% | 0 | 0 | — |
case-08 | fail→pass | 10,937 | 5,500 | -50% | 1 | 1 | 0% | 2,159 | 3,453 | +60% | 0 | 0 | — |
case-09 | fail→pass | 7,001 | 3,750 | -46% | 1 | 1 | 0% | 1,156 | 3,104 | +169% | 0 | 0 | — |
case-18 | pass→pass | 13,211 | 12,968 | -2% | 1 | 1 | 0% | 2,709 | 5,087 | +88% | 0 | 0 | — |
case-10 | pass→pass | 3,383 | 3,249 | -4% | 1 | 1 | 0% | 639 | 3,085 | +383% | 0 | 0 | — |
case-11 | fail→pass | 11,126 | 9,074 | -18% | 1 | 1 | 0% | 1,956 | 4,236 | +117% | 0 | 0 | — |
case-12 | pass→pass | 8,330 | 5,256 | -37% | 1 | 1 | 0% | 1,433 | 3,407 | +138% | 0 | 0 | — |
case-13 | fail→pass | 8,555 | 5,331 | -38% | 1 | 1 | 0% | 1,451 | 3,391 | +134% | 0 | 0 | — |
case-14 | fail→pass | 9,475 | 4,546 | -52% | 1 | 1 | 0% | 1,447 | 3,270 | +126% | 0 | 0 | — |
case-15 | fail→pass | 4,103 | 2,752 | -33% | 1 | 1 | 0% | 700 | 2,899 | +314% | 0 | 0 | — |
case-16 | fail→pass | 16,138 | 11,858 | -27% | 1 | 1 | 0% | 2,858 | 4,591 | +61% | 0 | 0 | — |
case-17 | pass→pass | 7,274 | 6,724 | -8% | 1 | 1 | 0% | 1,547 | 3,793 | +145% | 0 | 0 | — |
case-19 | pass→pass | 11,922 | 8,014 | -33% | 1 | 1 | 0% | 2,393 | 4,024 | +68% | 0 | 0 | — |
case-20 | fail→pass | 17,148 | 2,809 | -84% | 1 | 1 | 0% | 1,452 | 2,940 | +102% | 0 | 0 | — |
case-21 | pass→pass | 11,835 | 4,204 | -64% | 1 | 1 | 0% | 1,736 | 3,230 | +86% | 0 | 0 | — |
case-22 | pass→pass | 8,535 | 5,289 | -38% | 1 | 1 | 0% | 1,490 | 3,367 | +126% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +55 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.