content-retrieval（内容检索）

Content Retrieval Protocol

Token-efficient retrieval system using markdown-native section detection.

Summary

Core Principle: Never read a full file when partial content suffices.

Choose retrieval depth by need:

• Exists? → Grep (files_with_matches)
• Count? → Grep (count)
• Lookup? → Grep (-C:2)
• Overview? → Read (limit:40)
• Section? → Grep heading → Read (offset, limit)
• Full? → Read (justify first)

Quick Reference

Depth Levels

Decision Tree

What do I need?
├─ Does X exist? ──────────► L0: Grep files_with_matches
├─ How many X? ────────────► L1: Grep count
├─ What is X's value? ─────► L2: Grep with context
├─ What does X do? ────────► L3: Read limit:40
├─ How to use X for Y? ────► L4: Section extraction
└─ Implement X fully? ─────► L5: Full read (target only)

Tool Selection

Section Extraction

How Sections Work (Markdown-Native)

Sections are defined by headings and horizontal rules:

## Section A
Content...

---                    ← Section A ends here (horizontal rule)

## Section B           ← Or section ends at next same-level heading
Content...

No custom markers needed. Standard markdown structure.

Extraction Algorithm

1. Grep("^## Section Name$", file)     → line 15
2. Grep("^---|^## ", file, offset: 16) → next break at line 30
3. Read(file, offset: 15, limit: 15)   → lines 15-30

Section End Detection

A section ends at the first of:

1. --- (horizontal rule)
2. ## (same-level heading)
3. # (higher-level heading)
4. EOF (end of file)

Examples

# Extract Summary section
Grep("^## Summary$", file) → line 20
Grep("^---|^## ", file, offset: 21) → line 28
Read(file, offset: 20, limit: 8)

# Extract Quick Reference section
Grep("^## Quick Reference$", file) → line 30
Grep("^---|^## ", file, offset: 31) → line 55
Read(file, offset: 30, limit: 25)

Retrieval Protocol

Rule 1: Start at Lowest Depth

WRONG: Read entire SKILL-INDEX.md to find one skill
RIGHT: Grep("skill-name", SKILL-INDEX.md, -C:1)

WRONG: Read 5 skill files to understand what they do
RIGHT: Read each with limit:40 (frontmatter + summary)

WRONG: Read entire agent file to check if it has a skill
RIGHT: Grep("skills:.*skill-name", agent.md)

Rule 2: Index First, Files Second

1. Check index file (SKILL-INDEX.md, AGENT-INDEX.md)
2. Extract needed info via grep
3. Only read source file if index insufficient

Rule 3: Frontmatter Contains Metadata

All files have metadata in first 20 lines:

---
name: identifier
description: "one-line summary"
layer: 1-4
keywords: [searchable, terms]
---

To get metadata: Read(file, limit: 20)

Rule 4: Headings Define Sections

Standard markdown headings define extractable sections:

## Summary          ← Grep target
Brief description.

---

## Quick Reference  ← Grep target
| Col | Col |

To extract section: Grep heading → find end → Read range

Rule 5: Full Read Only for Targets

Full read justified when:

• File is the TARGET being modified/implemented
• L3 read confirmed full content needed
• Creating comprehensive documentation
• Debugging requires full context

QMD File Structure

Standard Layout

---
name: identifier
description: "One-line (< 100 chars)"
layer: 1-4
keywords: [grep, targets]
---

# Title

## Summary

2-3 sentences. Key purpose.

---

## Quick Reference

| Pattern | Usage |
|---------|-------|

---

## Patterns

### Pattern 1
Content...

### Pattern 2
Content...

---

## Related

- [link](path)

Reserved Headings

Frontmatter Schema

---
# Required
name: string           # Identifier (grep target)
description: string    # One-line (< 100 chars)
layer: number          # 1=foundation, 2=framework, 3=feature, 4=workflow

# Recommended
keywords: string[]     # Grep targets
depends_on: string[]   # Prerequisites
complements: string[]  # Often-used-with

# Optional
used_by: string[]      # Agents/commands using this
tech_stack: string[]   # Technologies
auto_apply: boolean    # Auto-trigger on match
---

Index File Format

Grep-Optimized Tables

## Master Index

| Name | L | Keywords | Path |
|------|---|----------|------|
| skill-a | 2 | key1,key2 | skills/skill-a/SKILL.md |
| skill-b | 1 | key3,key4 | skills/skill-b/SKILL.md |

---

## Lookup by Keyword

| Keyword | Skills |
|---------|--------|
| entity | abp-framework, abp-entity, domain-modeling |

Grep usage:

Grep("skill-a", INDEX.md, output: content, -C: 0)
→ | skill-a | 2 | key1,key2 | skills/skill-a/SKILL.md |

Examples

Example 1: Analyze Agent

Old approach (2,434 lines):

Read(AGENT-QUICK-REF.md)        → 127 lines
Read(abp-developer.md)          → 144 lines
Read(SKILL-INDEX.md)            → 337 lines
Read(3 skill files)             → 1,283 lines

Protocol approach (~265 lines):

Read(abp-developer.md)                    → 108 lines (target)
Grep("abp-developer", AGENT-INDEX.md)     → 1 line
Grep(skills from agent, SKILL-INDEX.md)   → 11 lines
Read(3 skills, limit:40 each)             → 120 lines

Savings: 89%

Example 2: Check Skill Exists

# L0: Just check existence
Grep("^name: xunit-testing", .claude/skills, output: files_with_matches)
→ .claude/skills/xunit-testing-patterns/SKILL.md

Example 3: Extract Section

# L4: Section extraction (markdown-native)
Grep("^## Quick Reference$", skill.md) → line 30
Grep("^---|^## ", skill.md, offset: 31) → line 55
Read(skill.md, offset: 30, limit: 25)
→ Just the Quick Reference section

Checklist

For File Authors

• Frontmatter under 20 lines
• description is one line, under 100 chars
• keywords include grep targets
• ## Summary section present (2-3 sentences)
• ## Quick Reference has lookup tables
• --- separators between major sections
• No critical info buried deep

For Consumers

• Check index before reading files
• Use appropriate depth level (L0-L5)
• Start with L0-L2, escalate if needed
• Full read only for target files

Related

• knowledge-discovery - What skills to find
• SKILL-INDEX.md - Skill lookup tables
• AGENT-INDEX.md - Agent lookup tables
• GUIDELINES.md - Standards and protocols