🗂️ Context Manipulation

Achieving infinite memory in finite context windows

---

The Problem

AI has context limits:

• Claude: ~200K tokens
• ChatGPT: ~128K tokens
• Others: even less

Large projects blow through this instantly. A medium codebase? 500K+ tokens easy.

You hit the wall. AI forgets. Progress stops.

The Solution

Context manipulation: Achieving 95-97% effective compression without losing critical information.

Not by upgrading models. By working smarter with context.

---

Core Techniques

1. Hierarchical Context Compression

Layer your context like an onion:

LAYER_0: Meta-context (what are we doing overall?)
LAYER_1: Structure (how is this organized?)
LAYER_2: Active work (what are we doing RIGHT NOW?)
LAYER_3: Details (specifics for current task)

Example:

@ai CONTEXT_LAYERS:

META: Refactoring authentication system
STRUCTURE: src/auth/ (5 files), tests/auth/ (3 files)
ACTIVE: Working on src/auth/login.ts
DETAIL: Lines 42-67, fixing token expiry check

Current issue: [specific problem]

Result:

• Traditional: Paste entire 500-line file (3,000+ tokens)
• Hierarchical: Reference + specific section (150 tokens)
• Compression: 95%

2. Context Anchoring

Define once, reference infinitely.

@ai ANCHOR:AUTH_SYSTEM

System: JWT-based authentication
Stack: Node.js + Express + PostgreSQL
Files:
  - src/auth/login.ts (main logic)
  - src/auth/tokens.ts (JWT generation)
  - src/auth/middleware.ts (route protection)
Pattern: All errors throw AuthError class
Tests: tests/auth/*.test.ts

// Now reference it:
@ai Using ANCHOR:AUTH_SYSTEM
Fix the token refresh race condition

Result:

• Define once: 200 tokens
• Reference forever: 10 tokens each time
• Savings: Massive over repeated conversations

3. Diff-Based Context

Send changes, not full files.

@ai FILE: src/auth/login.ts

CHANGED_LINES:
42: - if (token.expires < Date.now()) {
42: + if (token.expires <= Date.now()) {
58: + // Add null check
59: + if (!user) return null;

Context: Fixing edge case in token expiry

Result:

• Full file: 3,000 tokens
• Diff only: 120 tokens
• Compression: 96%

4. Pointer-Based Context

Use file paths instead of content.

@ai ISSUE: Authentication failing in production

Relevant files:
- src/auth/login.ts:42-67 (token validation)
- src/auth/tokens.ts:23 (token generation)
- config/jwt.ts:8 (secret key config)

Error occurs at login.ts:58
Related change in commit abc123

Investigate and fix.

Result:

• Pasting all files: 5,000+ tokens
• Pointers: 100 tokens
• Compression: 98%

AI can ask for specific files if needed. Start minimal.

5. Session State Management

Explicitly track conversation state.

@ai SESSION_STATE:

TASK: Refactor authentication
PROGRESS: 60% complete
COMPLETED:
  ✅ Login endpoint refactored
  ✅ Token generation updated
  ✅ Tests passing
IN_PROGRESS:
  🔄 Refresh token logic
BLOCKED:
  ⚠️ Waiting on database migration
NEXT:
  📋 Update middleware
  📋 Add rate limiting

Current focus: Refresh token implementation

Result:

• AI knows exactly where we are
• No need to repeat completed work
• Clear what’s next

6. Semantic Compression Codes

Create shorthand for complex patterns.

@ai CODES:

REFACTOR_PATTERN_1: Extract to function, add types, write test
ERROR_PATTERN: Try-catch, log to service, return Result type
API_PATTERN: Validate input, check auth, call service, return JSON

// Usage:
@ai Apply REFACTOR_PATTERN_1 to calculateTotal() function
@ai Add ERROR_PATTERN to all database queries

Result:

• Define patterns once
• Reference in 2-3 tokens
• AI knows exactly what you mean

7. Conditional Context Loading

Load context only when needed.

@ai I'm working on authentication

// AI asks: "Do you need context for:"
// - JWT implementation?
// - Database schema?
// - API endpoints?
// - Test patterns?

// You answer:
Just JWT implementation and tests

// AI loads ONLY those contexts

Result:

• Don’t load what you don’t need
• AI asks for more if required
• Stay under limits

8. Context Stack (Push/Pop)

Manage context like a stack.

@ai PUSH_CONTEXT: Refactoring login.ts

[Work on login.ts]

@ai POP_CONTEXT
@ai PUSH_CONTEXT: Updating tests

[Work on tests]

@ai POP_CONTEXT
Back to overall refactoring

Result:

• Deep work without losing place
• Return to previous state cleanly
• Organized context management

9. External Memory Files

Store context outside conversation.

Create a context.md file:

# Project Context

## Architecture
- Microservices: auth, users, billing
- Tech: Node.js + PostgreSQL + Redis
- Deploy: Kubernetes on AWS

## Current Work
Working on: Authentication refactor
Branch: feature/auth-v2
Status: 60% complete

## Key Files
- src/auth/login.ts - Main login logic
- src/auth/tokens.ts - JWT handling
- tests/auth/ - Test suite

Usage:

@ai Read context.md, then help me fix the token refresh bug

Result:

• Context persists across sessions
• Update once, use everywhere
• AI reads from file, not conversation

10. Lossy Compression

Keep essence, regenerate details.

@ai COMPRESSED_CONTEXT:

Auth system broken in production
- Error: "Token invalid"
- Happens: 5% of logins
- Started: After deploy xyz
- Suspect: Token expiry logic

Fix it.

// AI can ask for:
// - Full error logs?
// - Relevant code?
// - Recent commits?

// You provide only what AI requests

Result:

• Start with minimum
• Add details on demand
• Never over-context

---

Real-World Example

Scenario: Debugging complex issue in 100K+ line codebase

Traditional Approach:

@ai Here's the entire codebase [paste]
ERROR: Context limit exceeded

Context Manipulation Approach:

@ai ANCHOR:PROJECT_CONTEXT
- Codebase: E-commerce platform
- Stack: React + Node.js + PostgreSQL
- Scale: 100K LOC, 500+ files
- Issue: Checkout failing for 5% of users

ACTIVE_INVESTIGATION:
- File: src/checkout/payment.ts:145
- Error: "Cannot read property 'id' of undefined"
- Started: 2 days ago after deploy
- Affected: Users with saved cards

RELEVANT_FILES (pointer-based):
- src/checkout/payment.ts:140-160
- src/models/PaymentMethod.ts:23-45
- src/services/stripe.ts:67

HYPOTHESIS:
Saved cards missing 'id' field in some edge case

DIFF from last working version:
payment.ts:145
- const cardId = card.id
+ const cardId = card?.stripeId || card.id

Help me debug and fix.

Token usage:

• Traditional: CRASH (over limit)
• Optimized: ~300 tokens
• Effective compression: 99%+

---

The Workflow

1. Define anchors for major systems/contexts
2. Use pointers instead of pasting code
3. Send diffs for changes
4. Layer context (meta → structure → active → detail)
5. Track session state explicitly
6. Load conditionally - only what’s needed
7. Store externally - context.md files
8. Compress lossily - essence first, details on demand

---

Best Practices

✅ DO:

• Start with minimum context
• Add detail only when AI asks
• Use file:line references
• Create anchors for recurring contexts
• Update context.md files
• Think in layers

❌ DON’T:

• Paste entire files unless required
• Repeat context AI already has
• Front-load every possible detail
• Ignore context structure

---

Measuring Success

Before context manipulation:

• Hit limits constantly
• Restart conversations frequently
• Lose progress
• Frustrated

After context manipulation:

• Rarely hit limits
• Conversations last indefinitely
• Continuous progress
• Productive

Compression ratios:

• Good: 80-90%
• Great: 90-95%
• Legendary: 95-99%

---

Tools That Help

Manual:

• context.md files in your project
• Anchor definitions at conversation start
• Explicit state tracking

Future (coming soon):

• IDE extensions for auto-compression
• Context managers
• Automatic anchor generation

---

Next Level

Combine with:

• Emoji Protocol - Compress even further
• Meta-Prompting - Generate optimal context structures
• Quantum Prompting - Context in superposition

---

Practice Exercise

Try this:

1. Take a complex project you’re working on
2. Create a context.md with project overview
3. Define 3 ANCHOR points for major systems
4. Next AI conversation: Use anchors + pointers only
5. Measure token savings

Goal: 90%+ compression on first try

---

“The best context is the minimum context that gets maximum results.”

Master context manipulation. Work in codebases 10x larger than your context window. 🗂️

---

← Back to Advanced Prompting