accelint-skill-manager（加速技能经理）

Skill Manager

NEVER Do When Creating Skills

• NEVER write tutorials explaining basics - Assume Claude knows standard concepts, libraries, and patterns. Focus on expert-only knowledge.
• NEVER put triggering information in body - "When to use" guidance belongs ONLY in the description field. The body is loaded after activation decision.
• NEVER dump everything in SKILL.md - Use progressive disclosure: core workflow in SKILL.md (<500 lines ideal), detailed content in references/, loaded on-demand.
• NEVER use generic warnings - "Be careful" and "avoid errors" are useless. Provide specific anti-patterns with concrete reasons.
• NEVER use same freedom level for all tasks - Creative domains (design, architecture) need high freedom with principles. Fragile operations (file formats, APIs) need low freedom with exact scripts.
• NEVER explain standard operations - Assume Claude knows how to read files, write code, use common libraries. Focus on non-obvious decisions and edge cases.
• NEVER include obvious procedures - "Step 1: Open file, Step 2: Edit, Step 3: Save" wastes tokens. Include only domain-specific workflows Claude wouldn't know.
• NEVER skip the anti-patterns section — It's half of expert knowledge. A skill without "NEVER Do" is missing what makes it valuable: the mistakes experts learned the hard way.
• NEVER write a vague description — "A skill for X" causes false positives and missed activations. The description must include concrete trigger phrases users actually say.
• NEVER mix creation and audit concerns — Creating a skill, refactoring a skill, and auditing a skill are distinct workflows. Each has different inputs, outputs, and success criteria.

Before Creating a Skill, Ask

Apply these tests to ensure the skill provides genuine value:

Knowledge Delta Test

• Does this capture what takes experts years to learn? If explaining basics or standard library usage, it's redundant.
• Am I explaining TO Claude or arming Claude? Skills should arm agents with expert knowledge, not teach them concepts.
• Is every paragraph earning its context space? Token economy matters - shared with system prompts, conversation history, and other skills.

Activation Economics

• Does the description answer WHAT, WHEN, and include KEYWORDS? Vague descriptions mean the skill never gets activated.
• Would an agent reading just the description know exactly when to use this? If unclear, the skill is invisible.

Freedom Calibration

• What happens if the agent makes a mistake? High consequence = low freedom (exact scripts). Low consequence = high freedom (principles).
• Is there one correct way or multiple valid approaches? One way = prescriptive. Multiple ways = guidance with examples.

Token Efficiency

• Can this be compressed without losing expert knowledge? References loaded on-demand save context.
• Are there repetitive procedures that could become scripts? Reusable code belongs in scripts/, not repeated in instructions.

How to Use

This skill uses progressive disclosure to minimize context usage:

1. Start with the Workflow (SKILL.md)

Follow the 4-step workflow below for skill creation or refactoring.

2. Reference Implementation Details (AGENTS.md)

Load AGENTS.md for file system conventions, naming patterns, and structural rules.

3. Load Specific References as Needed

Each workflow step below notes which reference files to load. Only load what you need for the current step:

• Directory structure → references/file-system.md
• Description field conventions → references/skill.md
• AGENTS.md patterns → references/agents.md
• Progressive disclosure rules → references/progressive-disclosure.md
• Reference file format → references/references.md
• Script conventions → references/scripts.md
• Asset guidelines → references/assets.md
• CHANGELOG maintenance → references/changelog.md

Do NOT load all references at once — load only the ones relevant to your current step.

Which Workflow Should You Follow?

Choose based on your task:

• Creating a new skill from scratch → Follow Skill Creation Workflow (Steps 1-4)
• Improving an existing skill → Jump to Step 4 (Edit the Skill)
• Auditing a skill for quality → Follow Skill Audit Workflow

Skill Creation Workflow

To create or refactor a skill, follow the "Skill Creation Workflow" in order, skipping steps only if there is a clear reason why they are not applicable.

Copy this checklist to track progress:

- [ ] Step 1: Understanding - Gather concrete examples of skill usage
- [ ] Step 2: Planning - Identify reusable scripts, references, assets
- [ ] Step 3: Initializing - Check existing skills, create directory structure
- [ ] Step 4: Editing - Write agent-focused content with procedural knowledge and update CHANGELOG

Include what rules from this skill are being applied, and why, in your summary.

Step 1: Understanding the Skill with Concrete Examples

Skip this step only when the skill's usage patterns are already clearly understood. It remains valuable even when working with an existing skill.

To create an effective skill, clearly understand concrete examples of how the skill will be used. This understanding can come from either direct user examples or generated examples that are validated with user feedback.

Example: Building an image-editor skill, ask:

• "What functionality? Editing, rotating, other?"
• "Usage examples?"
• "Trigger phrases: 'Remove red-eye', 'Rotate image'—others?"

Ask 2-3 concrete questions first (functionality, examples, trigger phrases), then follow up based on their answers rather than front-loading all questions.

Conclude when there is a clear sense of the functionality the skill should support.

Step 2: Planning the Reusable Skill Contents

To turn concrete examples into an effective skill, analyze each example by:

1. Considering how to execute on the example from scratch
2. Identifying what scripts, references, and assets would be helpful when executing these workflows repeatedly

Examples:

• pdf-editor skill for "Rotate this PDF" → store scripts/rotate-pdf.sh to avoid rewriting code each time
• frontend-app-builder for "Build a todo app" → store assets/hello-world/ boilerplate template
• big-query for "How many users logged in today?" → store references/schema.md with table schemas

Analyze each concrete example to create a list of reusable resources: scripts, references, and assets.

Step 3: Initializing the Skill

MANDATORY: Load references/file-system.md before creating directory structure.

For new skills: Copy the template in assets/skill-template/ as a starting point and customize it.

For existing skills being refactored: Skip directly to Step 4 — the skill already exists.

Before creating, check for existing skills that overlap:

ls -la .claude/skills 2>/dev/null || echo "No project skills found"
ls -la ~/.claude/skills 2>/dev/null || echo "No global skills found"

If relevant skills exist, mention them briefly: "I found [list] — should any of these be included or merged?"

Follow the conventions in AGENTS.md and reference files for directory structure and naming.

Step 4: Edit the Skill

MANDATORY: Load references/skill.md for description field conventions and frontmatter rules.

When editing the (newly-generated or existing) skill, remember that the skill is being created for another instance of an agent to use. Focus on including information that would be beneficial and non-obvious to an agent. Consider what procedural knowledge, domain-specific details, or reusable assets would help another agent instance execute these tasks more effectively.

Calibrate freedom to task fragility:

The test: "If the agent makes a mistake, what's the consequence?"

• High consequence (file corruption, data loss) → Low freedom with precise scripts
• Medium consequence (suboptimal code, style issues) → Medium freedom with examples
• Low consequence (aesthetic choices, multiple valid approaches) → High freedom with principles

If you are updating an existing skill you can use the templates in assets/skill-template/ as a reference for larger structural changes and alignment. Consistency is imperative so lean towards aggressive reformatting to achieve adherence.

When updating an existing skill, ensure that the frontmatter metadata.version value is bumped. If the scope of the change is substantial do a major change 1.0 to 2.0, otherwise minor 1.0 to 1.1.

Version Control:

• Major (1.0 → 2.0): Substantial rewrites, breaking changes, complete restructuring
• Minor (1.0 → 1.1): New sections, significant additions, refinements
• Patch (1.0.0 → 1.0.1): Bug fixes, typo corrections (optional third digit)

CHANGELOG Maintenance: After updating a skill, update or create CHANGELOG.md using "Keep a Changelog" format:

# Changelog

## [X.Y.Z] - YYYY-MM-DD

### Added
- New features/capabilities with rationale

### Changed
- Modifications with why (always include rationale)

### Fixed
- Bug fixes with explanation

### Version
- Version bump note

Document what changed and why — the rationale is critical for future maintainers. Link to evaluation results when improvements stem from testing. The CHANGELOG version must match the frontmatter metadata.version.

Skill Audit Workflow

When auditing or reviewing an existing skill (not creating from scratch), follow this structured approach:

1. Frontmatter Audit

Check each field against requirements:

• name: lowercase, hyphens only, ≤64 chars, matches directory name
• description: starts with "Use when", includes WHAT/WHEN/KEYWORDS, has concrete trigger phrases
• license: present (optional but recommended)
• metadata.version: present, meaningful, and matches CHANGELOG.md latest version

2. Structure Audit

Compare against expected sections: NEVER Do, Before [Action] Ask, How to Use, Main Workflow. Note missing sections.

3. CHANGELOG and Version Audit

Check CHANGELOG.md presence and quality:

• Missing CHANGELOG: Flag as missing documentation
• Version mismatch: Verify CHANGELOG latest version matches frontmatter metadata.version
• Missing rationale: Ensure changes include WHY, not just WHAT
• Format compliance: Check for proper "Keep a Changelog" structure
• Empty sections: Verify Added/Changed/Fixed sections have meaningful content

4. Knowledge Delta Test

For each content block, ask: "Does Claude already know this?" Mark as REDUNDANT or EXPERT-ONLY. Calculate the percentage of redundant content. If >50% redundant, recommend substantial revision.

5. Produce Actionable Output

• Provide specific improvement recommendations ranked by priority
• Include a concrete improved description (not just criticism)
• Provide an improved SKILL.md alongside the audit report