Claude Long Document Strategies: Structuring 100K+ Token Prompts
Master Claude's 200K context for massive documents. Learn where to place instructions in long prompts, chunking strategies, progressive disclosure, and maintaining coherence across entire codebases and book-length documents.
Claude's 200K token context window works — genuinely, measurably, across the full range. You can load an entire codebase, a book manuscript, or a year of customer support logs and Claude will reason across it. But the strategies that work for 4K prompts don't scale. Where you place instructions, how you structure long documents, and how you manage Claude's attention all change at 100K+ tokens.
The Position Problem
In short prompts, instructions go at the beginning and everything works. In 100K+ prompts, the beginning is 80,000 words ago by the time Claude reaches the relevant content. Instruction placement becomes a retrieval problem.
The Sandwich Pattern
[BEGINNING: 5% of tokens]
Core instructions, role definition, output format specification.
These are your "always active" directives that apply to the entire document.
[MIDDLE: 90% of tokens]
The long document(s) — codebase, book, dataset, conversation history.
Structure this section for Claude's retrieval patterns (see below).
[END: 5% of tokens]
Specific task instructions: "Given the document above, analyze..."
Repeat critical constraints: "Remember: output must be valid JSON..."
Include specific reference points: "See the section on [topic] around line [X]..."
This pattern works because Claude's attention follows a U-shape — strongest at the beginning and end of the context window.
The Bookend Technique
For very long documents, place navigational anchors:
BEGINNING ANCHOR:
"This document contains three sections:
1. Project Requirements (pages 1-50)
2. Technical Specification (pages 51-120)
3. Test Plans (pages 121-180)
Use these section markers to navigate to relevant content."
... [long document content with clear section headers] ...
END ANCHOR:
"You've just read the full technical specification.
Key sections to reference:
- Authentication flow: Section 2.3 (around "AuthService")
- Database schema: Section 2.7 (around "CREATE TABLE")
- API contracts: Section 2.12 (around "OpenAPI spec")
Focus your analysis on sections 2.3-2.12 unless the question specifies otherwise."
Document Structure for Retrieval
Claude finds information better when documents are clearly structured:
Strong Structure (Claude navigates well)
### SECTION 2.3: Authentication Flow
The authentication system uses JWT tokens with the following flow:
1. Client sends credentials to /auth/login
2. Server validates and returns access_token (15min) + refresh_token (7 days)
3. Client includes access_token in Authorization header
4. On 401, client uses refresh_token at /auth/refresh
Key files: src/auth/AuthService.ts, src/auth/TokenManager.ts
Database tables: users, refresh_tokens, auth_attempts
Weak Structure (Claude struggles)
So for auth we basically use JWTs and the client sends stuff to the server
and gets back tokens. There's an access token that's short-lived and a refresh
token too. The code is mostly in the auth folder somewhere. We had some issues
with token expiry but fixed them last sprint.
Note:
The header rule: Every logical section needs a descriptive, unique header. Claude uses headers as navigation beacons in long contexts. "Section 2.3: Authentication Flow" is findable. Vague headers make content effectively invisible.
Progressive Disclosure Strategies
Don't dump 200K tokens at once. Build context progressively:
Strategy 1: The Funnel
Turn 1: "Here's the project README and architecture overview. How would you approach [task]?"
Turn 2: "Good. Now here are the relevant source files: [paste files]. Refine your approach."
Turn 3: "Here are the test files for those modules: [paste tests]. Identify edge cases your approach misses."
Turn 4: "Final round. Here are the deployment configs: [paste configs]. What production concerns should we address?"
Each turn adds context, narrowing the focus. By turn 4, Claude has seen the full picture without any single turn exceeding its attention budget.
Strategy 2: Index-First
Turn 1: "Here's a directory tree of the entire codebase. Identify which files are relevant to [task]."
Turn 2: "Now load ONLY the files you identified: [Claude lists files]. Here they are: [paste only relevant files]."
Let Claude do the filtering. It's better at identifying relevant sections than you might expect.
Attention Management for Analysis Tasks
When asking Claude to analyze a long document:
"Analyze the attached document for [specific thing].
Approach:
1. First, scan the entire document and identify all sections that discuss [topic].
2. Focus your detailed analysis on those sections.
3. For sections NOT about [topic], only note if they contain surprising or contradictory information.
4. Provide your analysis organized by section, with page/line references.
If you find conflicting information within the document, flag it explicitly:
'CONFLICT: Section 2 says [X] but Section 7 says [Y].'"
Codebase Analysis Pattern
"I'm attaching the entire codebase. Your task: identify security vulnerabilities.
Scan methodology:
1. Start with authentication/authorization files
2. Then input validation and sanitization
3. Then data access and database queries
4. Then configuration and secret management
5. Then dependency versions (package.json, requirements.txt)
For each finding, reference the exact file and line number.
Ignore: formatting issues, style inconsistencies, performance optimizations.
Focus ONLY on security vulnerabilities.
Note:
Anti-pattern: "Here's a 150K document. What do you think?" Claude will give you a generic summary that misses specifics. Always provide a focused task with explicit scanning instructions for long documents.
Related Pages
- Needle-in-Haystack Patterns — Once your documents are well-structured, learn to find specific information reliably within massive contexts.
- Context Window Economics — Decide whether full-context loading is cost-effective for your volume. RAG vs. full context vs. summarization chains.
Related Articles
Optimization Techniques
Master optimization strategies with effective prompts and practical approaches for ChatGPT.
Midjourney Weapon Design: Master Prompts for Swords, Guns & More
Master Midjourney weapon design and creation. Learn to craft stunning swords, futuristic guns, and fantasy armaments with advanced prompting techniques and detailed examples.
Interview Preparation with ChatGPT Prompts
Master job interviews with AI-powered preparation strategies. Get expert prompts for behavioral questions, technical interviews, and follow-up tactics.