writing-documentation-for-llms

Writing Documentation for LLMs

Guidance for creating effective documentation and instructions that LLMs can discover, understand, and use successfully.

Contents

• Core Principles
• Structure & Progressive Disclosure
• Content Patterns
• Anti-patterns
• Testing & Iteration

---

Core Principles

Assume competence

The LLM is already very smart. Only add information the LLM doesn't have. Challenge every piece:

• "Does the LLM really need this explanation?"
• "Can I assume the LLM knows this?"
• "Does this justify its token cost?"

Verbose example (~150 tokens):

PDF (Portable Document Format) files are a common file format that contains
text, images, and other content. To extract text from a PDF, you'll need to
use a library. There are many libraries available for PDF processing...

Concise example (~50 tokens):

Use pdfplumber for text extraction:
import pdfplumber
with pdfplumber.open("file.pdf") as pdf:
    text = pdf.pages[0].extract_text()

Match specificity to task fragility

• Narrow instructions (low freedom): database migrations, destructive operations, consistency-critical tasks
• General guidance (high freedom): code reviews, design decisions, where context determines best path

Test across models

Effectiveness varies by model. Skills that work for Claude Opus may need more detail for Claude Haiku.

---

Structure & Progressive Disclosure

Organize like a table of contents

Main file provides overview and points to detailed materials. LLM reads additional files only when needed.

Pattern:

• Main file: high-level guide with references (< 500 lines)
• Reference files: one per domain or topic
• Keep references one level deep (avoid chains: A → B → C)

Example structure:

my-doc/
├── OVERVIEW.md           # High-level guide
├── reference/
│   ├── api.md           # Specific reference
│   ├── examples.md      # Usage examples
│   └── troubleshooting.md
└── scripts/
    └── helper.py        # Executable utilities

Table of contents for long files

For any file over 100 lines, include a TOC at the top so LLM sees full scope:

## Contents
- Authentication and setup
- Core methods (create, read, update, delete)
- Advanced features
- Error handling patterns

Consistent terminology

Choose one term per concept and use it throughout:

• Always "API endpoint" (not "URL", "route", "path")
• Always "field" (not "box", "element", "control")
• Always "extract" (not "pull", "get", "retrieve")

---

Content Patterns

Descriptions: what + when

Enable discovery with concrete descriptions:

1. What it does: The concrete capability
2. When to use it: Specific triggers and contexts

Good example:

Extract text and tables from PDF files, fill forms, merge documents.
Use when working with PDF files or when the user mentions PDFs, forms, or document extraction.

Bad example:

Helps with documents

Examples over explanations

Show concrete input/output before abstract descriptions:

## Generating commit messages

Follow these examples:

**Example 1:**
Input: Added user authentication with JWT tokens
Output:
feat(auth): implement JWT-based authentication

Add login endpoint and token validation middleware

Workflows with clear steps

Break complex operations into sequential steps:

## Database migration workflow

Task Progress:
- [ ] Step 1: Create backup
- [ ] Step 2: Run migration script
- [ ] Step 3: Verify schema
- [ ] Step 4: Validate data integrity

Feedback loops for critical tasks

Use validators for quality-critical work:

## Document editing process

1. Make your edits
2. **Validate**: Run `validate.py`
3. If validation fails:
   - Review errors
   - Fix issues
   - Run validation again
4. **Only proceed when validation passes**
5. Finalize output

Conditional workflows

Guide the LLM through decision points:

## Modification workflow

1. Determine the modification type:
   - Creating new content? → Follow "Creation workflow"
   - Editing existing content? → Follow "Editing workflow"

2. Creation workflow:
   - Use library X
   - Build from scratch
   - Export format Y

3. Editing workflow:
   - Unpack existing file
   - Modify content
   - Repack when complete

Verifiable intermediate outputs

For complex tasks, create verifiable intermediate formats:

## Batch update workflow

1. Create plan file (JSON format)
2. **Validate plan**: Run `validate_plan.py`
3. If validation passes, execute
4. Verify output matches plan

Avoid time-sensitive information

Use "old patterns" sections for deprecated approaches:

## Current method

Use the v2 API endpoint: `api.example.com/v2/messages`

## Old patterns

<details>
<summary>Legacy v1 API (deprecated 2025-08)</summary>
The v1 API used: `api.example.com/v1/messages`
</details>

---

Anti-patterns

Too many options

Don't present multiple approaches unless necessary.

Bad:

You can use pypdf, or pdfplumber, or PyMuPDF, or pdf2image, or...

Good:

Use pdfplumber for text extraction.

For scanned PDFs requiring OCR, use pdf2image with pytesseract.

Deeply nested references

Keep references one level deep. Nested chains (file A → file B → file C) cause partial reads and missed context.

Vague trigger terms

Be specific in descriptions for discovery:

Vague: Helps with data

Specific: Analyze Excel spreadsheets, generate pivot tables, create charts. Use when working with Excel files, spreadsheets, or .xlsx files.

Windows-style paths

Always use forward slashes (Unix style):

• ✓ Good: scripts/helper.py
• ✗ Wrong: scripts\helper.py

---

Testing & Iteration

Create evaluations first

Before writing extensive documentation:

1. Identify gaps with LLM working without docs
2. Create 3+ representative test cases
3. Establish baseline performance
4. Write minimal docs to address gaps
5. Test and iterate based on results

Develop iteratively with Claude

1. Complete a task with Claude without docs
2. Identify the reusable pattern
3. Ask Claude to create docs capturing that pattern
4. Review for conciseness (remove explanations Claude already knows)
5. Test docs with a fresh instance on similar tasks
6. Iterate based on observations

---

Key Takeaway

Effective LLM documentation assumes intelligence, uses examples over explanation, organizes for progressive discovery, and validates critical workflows. Test with target models and iterate based on real usage patterns.