Rule Extraction

Rule Extraction

This skill teaches you how to extract a structured, line-numbered list of rules from a skill directory.

What Counts as a Rule

A rule is any statement that prescribes or constrains agent behavior. Recognize rules by their form:

• Imperative directives: "Always add type annotations", "Never use var"
• Prohibition patterns: "Do not X", "Avoid X", "禁止 X"
• Requirement patterns: "Must X", "Should X", "需要 X", "必须 X"
• Conditional behavior: "If X, then Y", "When X, do Y"
• Style/format mandates: "Use camelCase", "Limit lines to 80 characters"
• Preference statements: "Prefer X over Y", "优先使用 X"

Rules can appear as:

• Bullet list items (- or * or +)
• Numbered list items (1., 2.)
• Bold or emphasized sentences (Always do X)
• Plain sentences within a paragraph that contain directive language
• Section headers that are themselves imperatives ("Never commit secrets")

What Does NOT Count as a Rule

Exclude the following from extraction:

• Descriptive text: Explanations of why a rule exists, background context
• Examples: Code blocks, sample outputs, "for example..." passages
• Metadata: YAML frontmatter, version info, author notes
• Headings that introduce a section (unless the heading itself is an imperative rule)
• Vague non-actionable statements: "Do good work", "Be helpful" — extract these but flag them as vague: true

Output Format

Return a JSON array. Each rule is an object:

[
  {
    "id": 1,
    "file": "SKILL.md",
    "line": 12,
    "text": "Always add type annotations to function parameters",
    "source_text": "- Always add type annotations to function parameters",
    "vague": false
  },
  {
    "id": 2,
    "file": "SKILL.md",
    "line": 15,
    "text": "Do not use var, prefer const or let",
    "source_text": "**Do not use `var`** — prefer `const` or `let`",
    "vague": false
  },
  {
    "id": 3,
    "file": "examples.md",
    "line": 4,
    "text": "Do good work",
    "source_text": "- Do good work",
    "vague": true
  }
]

Field definitions:

• id: Sequential integer starting from 1, across all files combined
• file: Filename (basename only, e.g. SKILL.md, examples.md) where the rule appears
• line: Line number within that file where the rule appears (1-indexed)
• text: Cleaned rule text, stripped of markdown formatting
• source_text: Original raw line as it appears in the file
• vague: true if the rule is too broad to be actionable

Input Scope

The input is a skill directory path (e.g. ~/.claude/skills/humanizer/).

Files to read: all .md files directly inside the skill directory — SKILL.md first, then any other .md files in alphabetical order.

Files to skip:

• Subdirectories and any files inside them (no recursive scan)
• Non-markdown files (.sh, .py, .js, .json, etc.)

Extraction Process

1. List all .md files at the top level of the skill directory (non-recursive). Read SKILL.md first if present; read remaining .md files in alphabetical order.
2. For each file, read it line by line, tracking line numbers within that file.
3. Skip frontmatter (everything between --- delimiters at the top of each file)
4. Skip code blocks (between ` fences) — these are examples, not rules
5. For each line, determine if it contains a rule using the criteria above
6. For paragraph text: split into sentences, check each sentence independently
7. Clean the extracted text: remove markdown syntax (-, *, **, backticks) but preserve the meaning
8. Assign line numbers based on where the rule text starts within its file
9. Record the source filename for every rule

Handling Ambiguous Cases

• Compound bullet: "Use camelCase and limit lines to 80 chars" → split into two rules, both on the same line number
• Nested bullets: Treat each nested item as an independent rule
• Conditional rules: Keep the full conditional ("If Python, use type hints") as one rule — do not split on the condition
• Negated rules: "Don't use X" and "Never use X" are equivalent — normalize to a consistent form in text, preserve original in source_text

Example

Input skill directory: ~/.claude/skills/code-style/ Files found at top level: SKILL.md, extra-rules.md (Any subdirectory contents are ignored.)

SKILL.md:

---
name: Code Style
---

# Code Style Guidelines

Keep code clean and maintainable.

## Naming
- Use camelCase for variables and functions
- Use PascalCase for classes and types
- Never use single-letter variable names except for loop indices

## Comments
Always write comments in English.
Do not write comments explaining what the code does — only explain why.

## Example
\`\`\`python
# good
def calculate_total(items):
    return sum(items)
\`\`\`

extra-rules.md:

- Prefer explicit returns over implicit ones

Output:

[
  {
    "id": 1,
    "file": "SKILL.md",
    "line": 10,
    "text": "Use camelCase for variables and functions",
    "source_text": "- Use camelCase for variables and functions",
    "vague": false
  },
  {
    "id": 2,
    "file": "SKILL.md",
    "line": 11,
    "text": "Use PascalCase for classes and types",
    "source_text": "- Use PascalCase for classes and types",
    "vague": false
  },
  {
    "id": 3,
    "file": "SKILL.md",
    "line": 12,
    "text": "Never use single-letter variable names except for loop indices",
    "source_text": "- Never use single-letter variable names except for loop indices",
    "vague": false
  },
  {
    "id": 4,
    "file": "SKILL.md",
    "line": 15,
    "text": "Always write comments in English",
    "source_text": "Always write comments in English.",
    "vague": false
  },
  {
    "id": 5,
    "file": "SKILL.md",
    "line": 16,
    "text": "Do not write comments explaining what the code does — only explain why",
    "source_text": "Do not write comments explaining what the code does — only explain why.",
    "vague": false
  },
  {
    "id": 6,
    "file": "extra-rules.md",
    "line": 1,
    "text": "Prefer explicit returns over implicit ones",
    "source_text": "- Prefer explicit returns over implicit ones",
    "vague": false
  }
]

Notes:

• SKILL.md lines 18–23 (the code block) are skipped entirely
• id is sequential across all files; line resets to 1 for each new file
• Any subdirectories inside ~/.claude/skills/code-style/ are not scanned