Documentation Navigation Skill

Navigate the hierarchical documentation system to find relevant context with minimal token usage.

Variables

Variable	Default	Description
START_LEVEL	auto	auto, root, category, library, section, page
KEYWORD_MATCH	fuzzy	fuzzy (semantic) or exact matching for index navigation

Instructions

MANDATORY - Always start at the appropriate index level and drill down.

• Read _index.toon files before loading pages
• Match task keywords against index keywords
• Load specific pages, not full contexts
• Note priority and when_to_use fields in indexes

Red Flags - STOP and Reconsider

If you're about to:

• Load full-context.md for a simple question
• Guess file paths instead of reading the index
• Skip the index and go straight to pages
• Load everything "just in case"

STOP -> Follow the Navigation Protocol below -> Then proceed

Workflow

1. CHECKPOINT: Identify what you know (library? section? topic?)
2. Start at the appropriate index level
3. Read index, match keywords to your task
4. Drill down to specific pages
5. Load only the pages you need
6. CHECKPOINT: Is full-context.md justified? (rarely)

Cookbook

Level Selection

• IF: Unsure where to start in the hierarchy
• THEN: Read cookbook/level-selection.md
• RESULT: Start at the right level based on what you know

Keyword Matching

• IF: Searching for documentation by topic
• THEN: Read cookbook/keyword-matching.md
• RESULT: Find relevant pages through index keywords

Efficient Page Loading

• IF: Loading multiple pages or managing token budget
• THEN: Read cookbook/page-loading.md
• RESULT: Minimize token usage while getting needed context

The Hierarchy

ai-docs/
├── _root_index.toon              # Level 0: All categories
├── libraries/
│   ├── _index.toon               # Level 1: All libraries
│   ├── {library}/
│   │   ├── _index.toon           # Level 2: Library sections
│   │   ├── {section}/
│   │   │   ├── _index.toon       # Level 3: Section pages
│   │   │   └── pages/
│   │   │       └── {page}.toon   # Level 4: Page summary
│   │   └── full-context.md       # Everything (escape hatch)

Navigation Protocol

Step 1: Start at the Right Level

You Know	Start Here
Nothing	ai-docs/_root_index.toon
Category (libraries/frameworks)	ai-docs/{category}/_index.toon
Library name	ai-docs/libraries/{lib}/_index.toon
Library + section	ai-docs/libraries/{lib}/{section}/_index.toon
Exact page	ai-docs/libraries/{lib}/{section}/pages/{page}.toon

Step 2: Read Index, Match Keywords

Each index contains:

• description or purpose - What this level contains
• keywords - Terms for matching your task
• when_to_use - Guidance on when to drill deeper
• Child paths - Where to navigate next

Match your task against keywords:

Task: "Handle LLM retry failures"
Keywords to find: retry, error, failure, fallback, timeout

Navigate:
1. libraries/_index.toon -> "baml" has "llm|structured-output"
2. baml/_index.toon -> common_tasks mentions "error handling"
3. baml/guide/_index.toon -> "error-handling" page has "retry|fallback"
4. baml/guide/pages/error-handling.toon -> Exact info needed

Step 3: Load Only What's Needed

DO:

@ai-docs/libraries/baml/guide/pages/error-handling.toon  # ~400 tokens

DON'T:

@ai-docs/libraries/baml/full-context.md  # ~15,000 tokens

Step 4: Load Multiple Pages if Needed

For complex tasks, load several specific pages:

@ai-docs/libraries/baml/guide/pages/error-handling.toon
@ai-docs/libraries/baml/guide/pages/timeouts.toon
@ai-docs/libraries/baml/reference/pages/retry-policy.toon

Still far cheaper than full-context.md.

Step 5: Use full-context.md Sparingly

Only for:

• Writing comprehensive tutorials
• Major migrations
• Initial deep learning of a library

Index File Structure

Category Index (_index.toon)

meta:
  level: category
  name: libraries
  count: 5

items[5]{id,name,description,path,keywords,priority}:
baml,BAML,Structured LLM outputs with type safety,./baml/_index.toon,llm|types|prompts,high
mcp,MCP,Tool integration for AI assistants,./mcp/_index.toon,tools|servers|context,high

Library Index ({lib}/_index.toon)

meta:
  level: library
  name: BAML
  version: 0.70.0
  homepage: https://docs.boundaryml.com

overview: |
  Brief description of what this library does and when to use it.

sections[3]{id,name,path,page_count,when_to_use}:
guide,Guide,./guide/_index.toon,15,Learning concepts and tutorials
reference,Reference,./reference/_index.toon,25,Exact syntax and API details
examples,Examples,./examples/_index.toon,8,Working code samples

common_tasks[5]{task,section,page}:
Define output types,reference,types
Handle errors,guide,error-handling
Stream responses,guide,streaming

Section Index ({lib}/{section}/_index.toon)

meta:
  level: section
  library: baml
  section: guide
  page_count: 15

pages[15]{id,title,path,priority,purpose,keywords}:
introduction,Introduction,./pages/introduction.toon,core,Overview and basic concepts,basics|overview|start
error-handling,Error Handling,./pages/error-handling.toon,important,Retry and fallback strategies,retry|error|fallback

Page Summary ({lib}/{section}/pages/{page}.toon)

meta:
  level: page
  library: baml
  section: guide
  page: error-handling
  source_url: https://docs.boundaryml.com/guide/error-handling
  content_hash: abc123

summary:
  purpose: |
    One to two sentences on what this page teaches.

  key_concepts[5]: concept1|concept2|concept3|concept4|concept5

  gotchas[3]: warning1|warning2|warning3

code_patterns[2]:
  - lang: baml
    desc: What this pattern shows
    code: |
      // Truncated code example

related_pages[2]: ../pages/timeouts.toon|../../reference/pages/retry-policy.toon

Navigation Examples

Example 1: Specific Feature

Task: "Configure BAML streaming"

1. @ai-docs/libraries/baml/_index.toon
   -> common_tasks: "Stream responses" -> guide/streaming

2. @ai-docs/libraries/baml/guide/pages/streaming.toon
   -> Get streaming configuration

Tokens: ~600

Example 2: Unknown Library

Task: "Type-safe database access in TypeScript"

1. @ai-docs/libraries/_index.toon
   -> Scan keywords: "prisma" has "database|orm|typescript"

2. @ai-docs/libraries/prisma/_index.toon
   -> Navigate to relevant section

Tokens: ~400 + next level

Example 3: Comprehensive Knowledge

Task: "Migrate entire codebase from X to BAML"

1. @ai-docs/libraries/baml/full-context.md
   -> Load everything (correct choice here)

Tokens: ~15,000 (justified for migration)

Anti-Patterns

Bad	Good	Why
Load full-context.md for simple question	Navigate to specific page	95% token waste
Guess at file paths	Start at index, drill down	Indexes show what exists
Load everything "just in case"	Load what task requires	Wastes context window
Skip index, go straight to pages	Read index first	May miss better pages