skill-creator（技能创建器）

Agent Skill Creator Standard

Priority: P0 (CRITICAL)

Strict guidelines for authoring High-Density Agent Skills. Maximize information density while minimizing token consumption through progressive disclosure and strategic content organization.

Core Principles

Token Economy First ⚡

Every word costs tokens. Design skills for maximum information/token ratio:

• Progressive Loading: Load only essential content initially
• Lazy References: Move detailed examples to references/ folder
• Imperative Compression: Use verbs, abbreviations, bullet points
• Context Window Awareness: Design for 4k-32k token limits across agents

Three-Level Loading System

Level 1: Metadata (100 words) → Always loaded
Level 2: SKILL.md Body (100 lines) → When triggered
Level 3: References/Scripts/Assets → As needed

Directory Structure

skills/
└── {category}/                     # e.g., "flutter" (lowercase)
    └── {skill-name}/               # e.g., "bloc-state-management" (kebab-case)
        ├── SKILL.md                # Core Logic (High Density, <100 lines)
        ├── scripts/                # Executable code (Deterministic tasks)
        │   └── automation.py
        ├── references/             # Heavy Examples (Lazy loaded)
        │   ├── patterns.md
        │   └── examples.md
        └── assets/                 # Output templates (Never loaded)
            └── template.json

Writing Rules (Token-Optimized)

1. Imperative Compression: Start with verbs. No "Please/You should".

- *Waste*: "You should use BLoC for state management." (8 words) - *Efficient*: "Use BLoC for state management." (5 words)

1. Token Economy: Maximize info/token ratio.

- Skip articles ("the", "a") if readable - Use standard abbreviations (cfg, param, impl) - Bullet points > paragraphs (3x density)

1. Progressive Disclosure: Essential info first, details on-demand.

- Core workflow in SKILL.md - Complex examples in references/ - Templates/assets never loaded

1. Context-Aware Design: Different agents have different limits.

- Cursor: ~100k tokens - Claude: ~200k tokens - Windsurf: ~32k tokens

Content Sections (Token-Budgeted)

Required sections in SKILL.md:

1. Frontmatter (Mandatory): Metadata for triggering (100 words max) --- name: Skill Name description: What it does + when to use it (triggers activation) metadata: labels: [tag1, tag2] triggers: files: ['**/*.ext'] keywords: [term1, term2] ---
2. Priority: P0 (Critical), P1 (Standard), or P2 (Optional)
3. Structure: ASCII tree of expected file layout
4. Guidelines: Bullet points of "Do this" (imperative)
5. Anti-Patterns: Bullet points of "Don't do this"
6. Reference Links: Links to references/ files (lazy loading)

Size Limits (Strict)

Resource Organization (Token-Saving)

scripts/ - Executable Code

When to use: Deterministic, repeated tasks Benefits: Never loaded into context, executed directly Examples: Code generators, formatters, validators

references/ - Documentation

When to use: Detailed examples, API docs, complex patterns Benefits: Loaded only when needed, keeps SKILL.md lean Examples: Implementation patterns, error handling guides

assets/ - Output Templates

When to use: Boilerplate files, images, configs Benefits: Never loaded, copied to output as-needed Examples: Project templates, config files, icons

Skill Creation Lifecycle

Phase 1: Understanding (Token Audit)

1. Define concrete use cases
2. Identify repetitive patterns
3. Calculate token budget per agent

Phase 2: Planning (Resource Strategy)

1. Core workflow → SKILL.md
2. Complex examples → references/
3. Repeated code → scripts/
4. Templates → assets/

Phase 3: Implementation (Compression)

1. Write imperative guidelines
2. Compress examples to essentials
3. Test context window fit

Phase 4: Validation (Token Testing)

1. Verify loading efficiency
2. Test across different agents
3. Measure token consumption

Validation Checklist

Before finalizing, verify:

• SKILL.md ≤100 lines (ideal: 40-60)
• No inline code >10 lines
• No repeated frontmatter content
• Anti-patterns use strict format (see below)
• Complex examples in references/
• Tables >8 rows moved to references/
• No description redundancy after Priority

Anti-Patterns (Token Wasters)

• Verbose Explanations: "This is important because..." → Delete
• Redundant Context: Same info in multiple places
• Large Inline Code: Move code >10 lines to references/
• Conversational Style: "Let's see how to..." → "Do this:"
• Over-Engineering: Complex structure for simple skills
• Redundant Descriptions: Do not repeat frontmatter description after ## Priority
• Oversized Skills: SKILL.md >100 lines → Extract to references/
• Nested Formatting: Avoid **Bold**: \More Bold`` - causes visual noise
• Verbose Anti-Patterns: See strict format below

Anti-Pattern Format (Strict)

Format: **No X**: Do Y[, not Z]. [Optional context, max 15 words total]

Examples:

❌ Verbose (24 words):

- **No Manual Emit**: `**Avoid .then()**: Do not call emit() inside Future.then; always use await or emit.forEach.`

✅ Compressed (11 words):

- **No .then()**: Use `await` or `emit.forEach()` to emit states.

❌ Verbose (18 words):

- **No UI Logic**: `**Logic in Builder**: Do not perform calculations or data formatting inside BlocBuilder.`

✅ Compressed (9 words):

- **No Logic in Builder**: Perform calculations in BLoC, not UI.

Progressive Disclosure Checklist

Extract to references/ when:

• Code examples >10 lines
• Tables >8 rows
• Explanatory sections >10 lines
• Multiple code variants/alternatives
• Detailed performance benchmarks
• Step-by-step tutorials

Reference & Examples

Use the enhanced template below to generate new skills: references/TEMPLATE.md

For comprehensive lifecycle guidance: references/lifecycle.md

For resource organization patterns: references/resource-organization.md