llm-boost（LLM boost 搜索）

LLM Boost Skill

Optimize all LLM-facing content: documentation, skills, prompts, and parameters.

Quick Reference

Documentation Optimization (c7score)

1. Analyze: Read README.md, docs/*.md
2. Generate questions: Create 15-20 "How do I..." questions covering setup, auth, basic usage, errors, advanced features, integrations
3. Map questions to snippets: Mark complete, partial, or missing (prioritize missing)
4. Optimize by priority:

1. Validate each snippet: runs standalone, answers specific question, proper format, includes imports
2. Score before vs after across all 5 metrics

Snippet Transformation Patterns

• API ref to usage example: Replace method signatures with complete working code including imports, setup, and expected output
• Import-only to complete setup: Combine from lib import X with actual usage showing real output
• Multiple fragments to one comprehensive: Merge related 1-2 line snippets into one complete workflow
• Remove metadata: Strip directory trees, license text, BibTeX citations entirely

For detailed patterns: references/optimization_patterns.md

llms.txt Generation

<format_rules>

• H1 title required, H2 sections only (no H3+)
• Full URLs with protocol, prefer.md files
• - [Title](url): description link format
• "Optional" section = skippable for shorter context
• No code blocks, images, or complex formatting
• Place at repo root as /llms.txt </format_rules>

For templates: examples/sample_llmstxt.md

Skill Optimization

500-Line Rule

Keep in SKILL.md: purpose, quick start, critical practices, brief examples (5-10 lines), cross-references. Move to reference files: API docs, extensive examples (>20 lines), troubleshooting, pattern libraries, schemas.

Optimization Modes

YAML Frontmatter

Description field (max 1024 chars) must include: what the skill does, when to use it, key technologies, action verbs. Write in third person.

Progressive Disclosure Pattern

## Topic Overview
Brief explanation (2-3 sentences).

**Quick Example:**
(5-10 line code block)

**For detailed docs**: [REFERENCE.md](REFERENCE.md#topic)

XML Tag Structuring

<design_principles>

</design_principles>

Core Patterns

Multi-document: <documents><document index="1"><source>...</source><content>...</content></document></documents>

Structured evaluation: <rubric> + <submission> -> <evaluation><score> + <feedback>

CoT separation: <thinking> for reasoning, <answer> for final output

Multishot examples: <examples><example><input>...</input><output>...</output></example></examples>

Guard rails: <instructions><task>...</task><formatting>...</formatting><constraints>...</constraints></instructions>

Output Extraction

import re

def extract_tag(text, tag):
    match = re.search(f'<{tag}>(.*?)</{tag}>', text, re.DOTALL)
    return match.group(1).strip() if match else None

For comprehensive tag catalog: references/xml_tags.md

LLM Parameter Tuning

CLAUDE.md Audit Checklist

Reference Materials

• c7score Metrics - scoring rubrics and weights
• Optimization Patterns - snippet transformation patterns
• llms.txt Format - complete format specification
• XML Tag Patterns - comprehensive tag catalog
• Skill Optimization - 3-level loading, migration workflow