llm-boost

What this skill does

# LLM Boost Skill

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

## Quick Reference

| Area       | Key Metric                | Target           |
| ---------- | ------------------------- | ---------------- |
| c7score    | Question-Snippet Match    | 80% weight       |
| Skills     | SKILL.md size             | <=500 lines      |
| LLM Tuning | Task-appropriate settings | See tuning table |

---

## Documentation Optimization (c7score)

<workflow>

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:

| Priority              | Weight | Action                                       |
| --------------------- | ------ | -------------------------------------------- |
| P1: Question coverage | 80%    | Add complete code for unanswered questions   |
| P2: Remove duplicates | 5%     | Consolidate similar snippets                 |
| P3: Fix formatting    | 5%     | Proper language tags, TITLE/DESCRIPTION/CODE |
| P4: Remove metadata   | 5%     | Strip licensing, directory trees, citations  |
| P5: Enhance init      | 5%     | Combine import-only with usage examples      |

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

</workflow>

### 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](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>

| Project Type | Must Have                                      | Should Have                  |
| ------------ | ---------------------------------------------- | ---------------------------- |
| Library      | Documentation, API Reference, Examples         | Getting Started, Development |
| CLI Tool     | Getting Started, Commands, Examples            | Configuration, Development   |
| Framework    | Documentation, Guides, API Reference, Examples | Integrations                 |

For templates: [examples/sample_llmstxt.md](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

| Mode       | Size       | Action                                      |
| ---------- | ---------- | ------------------------------------------- |
| Light      | <3K tokens | Tighten wording, add YAML if missing        |
| Standard   | 3K-6K      | Consolidate, tables over prose, one example |
| Aggressive | 6K-10K     | Table everything, strip filler              |
| Split      | >=10K      | Propose 3-4 files + index                   |

### 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

```markdown
## 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>

| Principle          | Guideline                                                            |
| ------------------ | -------------------------------------------------------------------- |
| Semantic naming    | Tag names describe content: `<contract>`, `<rubric>`                 |
| Consistency        | Same tag names throughout; reference by name in instructions         |
| Nesting            | `<outer><inner></inner></outer>` for hierarchy                       |
| No canonical tags  | No "best" tags - name for your use case                              |
| Combine techniques | Pair with CoT (`<thinking>`/`<answer>`) and multishot (`<examples>`) |

</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

```python
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](references/xml_tags.md)

---

## LLM Parameter Tuning

| Task                 | max_tokens | temperature | top_p | Rationale                                     |
| -------------------- | ---------- | ----------- | ----- | --------------------------------------------- |
| Theorem proving      | 4096       | 0.6         | 0.95  | CoT needs space; higher temp explores tactics |
| Code generation      | 2048       | 0.2-0.4     | -     | Deterministic preferred                       |
| Creative/exploration | 4096       | 0.8-1.0     | -     | Maximum diversity                             |
| Classification       | 256        | 0.0-0.1     | -     | Consistency over creativity                   |
| Summarization        | 1024       | 0.3         | -     | Faithful to source                            |

---

## CLAUDE.md Audit Checklist

| Check                  | How                                    |
| ---------------------- | -------------------------------------- |
| Tech stack claims      | `Read("package.json\|Cargo.toml")`     |
| File path references   | `Glob("claimed/path")`                 |
| Command references     | `Grep("script", glob="package.json")`  |
| Testing framework      | `Glob("**/*.test.*")`                  |
| Linting config         | `Glob("**/biome.json\|**/.eslintrc*")` |
| Line count             | `wc -l CLAUDE.md` - target <300        |
| No code duplication    | Uses file:line pointers                |
| WHAT/WHY/HOW structure | Manual review                          |

---

## Reference Materials

- [c7score Metrics](references/c7score_metrics.md) - scoring rubrics and weights
- [Optimization Patterns](references/optimization_patterns.md) - snippet transformation patterns
- [llms.txt Format](references/llmstxt_format.md) - complete format specification
- [XML Tag Patterns](references/xml_tags.md) - comprehensive tag catalog
- [Skill Optimization](references/skill_optimization.md) - 3-level loading, migration workflow