enhance-docs

enhance-docs

Analyze documentation for readability, structure, and RAG optimization.

Parse Arguments

const args = '$ARGUMENTS'.split(' ').filter(Boolean);
const targetPath = args.find(a => !a.startsWith('--')) || '.';
const fix = args.includes('--fix');
const aiMode = args.includes('--ai');

Documentation Locations

Type	Location	Purpose
User docs	docs/*.md, README.md	Human-readable guides
Agent docs	agent-docs/*.md	AI reference material
Project memory	CLAUDE.md, AGENTS.md	AI context/instructions

Optimization Modes

AI-Only Mode (--ai)

For agent-docs and RAG-optimized documentation:

• Aggressive token reduction
• Dense information packing
• Self-contained sections for retrieval
• Optimal chunking boundaries

Both Mode (--both, default)

For user-facing documentation:

• Balance readability with AI-friendliness
• Clear structure for both humans and retrievers

Workflow

1. Discover - Find all .md files
2. Parse - Extract structure and content
3. Check - Run pattern checks based on mode
4. Report - Generate markdown output
5. Fix - Apply auto-fixes if --fix

Detection Patterns

1. Link Validation (HIGH)

• Broken anchor links ([text](#missing-anchor))
• Links to non-existent files
• Malformed link syntax

2. Structure Validation (HIGH)

Heading hierarchy:

• No jumps (H1 → H3 without H2)
• Single H1 per document
• Code blocks with language tags

Position-aware content (based on "lost in the middle" research):

• Critical info at START or END of document
• Supporting details in MIDDLE
• Flag important content buried in middle sections

Recommended structure:

1. Overview/Purpose (START - high attention)
2. Quick Start / TL;DR
3. Detailed Content
4. Reference / API
5. Summary / Key Points (END - high attention)

3. Token Efficiency (HIGH - AI Mode)

Token estimation: characters / 4 or words * 1.3

Unnecessary prose:

• "In this document..."
• "As you can see..."
• "Let's explore..."
• "It's important to note that..."

Verbose phrases:

Verbose	Concise
"in order to"	"to"
"due to the fact that"	"because"
"has the ability to"	"can"
"at this point in time"	"now"
"for the purpose of"	"for"
"in the event that"	"if"

Target: ~1500 tokens for project memory files, flexible for reference docs.

4. RAG Optimization (MEDIUM - AI Mode)

Chunk size guidelines:

Size	Issue
>1000 tokens	Too long, split into subtopics
<50 tokens	Too short, merge with related content
200-500 tokens	Optimal for retrieval

Semantic boundaries:

• Single topic per section
• Self-contained sections (avoid "It", "This" at section start)
• Clear section titles that describe content

Context anchors:

# Bad - ambiguous start
## Configuration
It requires several settings...

# Good - self-contained
## Configuration
The plugin configuration requires several settings...

5. Information Density (MEDIUM - AI Mode)

Prefer tables over prose:

# Bad - verbose
The function accepts a path parameter which is required,
a limit parameter which defaults to 10, and an optional
format parameter.

# Good - dense
| Param | Required | Default | Description |
|-------|----------|---------|-------------|
| path | Yes | - | File path |
| limit | No | 10 | Max results |
| format | No | json | Output format |

Prefer lists over paragraphs for sequential items.

Use code blocks for examples, commands, configurations.

6. Cross-Reference Quality (MEDIUM)

• Internal links should use relative paths
• External links should be stable (avoid commit hashes)
• Reference sections should point to canonical sources

7. Balance Suggestions (MEDIUM - Both Mode)

• Missing section headers in long content (>500 words without heading)
• Important information buried late in document
• Missing TL;DR or summary for long documents

Auto-Fixes

Issue	Fix
Inconsistent headings	H1 → H3 becomes H1 → H2
Verbose phrases	Replace with concise alternatives
Missing code language	Add based on content detection

Output Format

## Documentation Analysis: {name}

**File**: {path}
**Mode**: {AI-only | Both}
**Tokens**: ~{count}

| Certainty | Count |
|-----------|-------|
| HIGH | {n} |
| MEDIUM | {n} |

### Link Issues
| Line | Issue | Fix | Certainty |

### Structure Issues
| Line | Issue | Fix | Certainty |

### Efficiency Issues [AI mode]
| Line | Issue | Fix | Certainty |

### RAG Issues [AI mode]
| Line | Issue | Fix | Certainty |

Pattern Statistics

Category	Patterns	Mode	Certainty
Links	3	shared	HIGH
Structure	4	shared	HIGH
Token Efficiency	3	ai	HIGH
RAG Optimization	3	ai	MEDIUM
Information Density	2	ai	MEDIUM
Cross-Reference	2	shared	MEDIUM
Balance	3	both	MEDIUM
Total	20	-	-

### Verbose Phrase ```markdown In order to configure the plugin, you need to... ``` ```markdown To configure the plugin... ```

RAG Chunking

<bad_example>

## Installation
[2000+ tokens of mixed content covering install, config, and usage]

</bad_example> <good_example>

## Installation
[400 tokens - installation only]

## Configuration
[300 tokens - config only]

## Usage
[400 tokens - usage only]

</good_example>

Position-Aware Content

<bad_example>

## Introduction
[Long background...]

## History
[More context...]

## Critical Setup Steps
[Important info buried in middle]

</bad_example> <good_example>

## Quick Start (Critical)
[Important setup steps at START]

## Background
[Supporting context in middle]

## Reference
[Details...]

## Key Reminders
[Critical points repeated at END]

</good_example>

Tables vs Prose

<bad_example>

The API accepts three parameters. The first is `query` which is required.
The second is `limit` which defaults to 10. The third is `format`.

</bad_example> <good_example>

| Param | Required | Default |
|-------|----------|---------|
| query | Yes | - |
| limit | No | 10 |
| format | No | json |

</good_example>

References

• agent-docs/CONTEXT-OPTIMIZATION-REFERENCE.md - Token budgeting, position awareness, chunking
• agent-docs/PROMPT-ENGINEERING-REFERENCE.md - Structure, information density

Constraints

• Auto-fix only HIGH certainty issues
• Preserve original tone and style
• Balance AI optimization with human readability (default mode)
• Don't remove content, only restructure or condense