The Self-Augmentation Framework

Don't just "keep up" with AI releases. Build a system for getting better.

These three compound. Tools without domain expertise = slop. Domain expertise without tools = slow. Both without experimentation = stagnant.

• LLMs are stateless — they don't remember anything between API calls
• Everything lives in the context window: your prompt, system instructions, tool definitions, tool call history, tool results, and the model's own responses
• Agents are token guzzlers — both tool calls AND their results stay in context
• The engineering challenge: maximize useful information within a finite window

How Context Actually Works

Everything accumulates in one big messages array

// What the model sees every API call:
{
  "system": "You are a coding agent...",        // ~500 tokens
  "tools": [bash, read, write, edit, ...],    // ~2,000 tokens
  "messages": [
    {"user": "Fix the auth bug"},              // ~10 tokens
    {"assistant": tool_use(read_file)},         // ~50 tokens
    {"user": tool_result("...500 lines...")},  // ~2,000 tokens
    {"assistant": tool_use(read_file)},         // ~50 tokens
    {"user": tool_result("...300 lines...")},  // ~1,200 tokens
    // ... 15 more tool calls ...
    // Context is now 80% full of file contents
  ]
}

Every file you read, every command you run — it all stays in the window. This is why agents "forget" things on long tasks. The useful information gets drowned out.

Context Rot

The effective window is ~50-60% of the stated capacity

• Quality degrades as context fills — model attention becomes diffuse
• Critical details from early in the conversation get "forgotten"
• Newer information tends to dominate over older context
• This is why long sessions produce worse results than fresh ones

Fighting Context Rot

Four strategies to keep context clean

The best agent users are context hygienists.

Sub-Agents

Specialized Claude instances for focused work

Skills & Hooks

# skills/my-skill/SKILL.md
---
name: my-skill
description: When to trigger this
---

# Instructions for the model
Step 1: ...
Step 2: ...

// .claude/settings.json
{
  "hooks": {
    "Stop": [{
      "command": "afplay /System/Library/
        Sounds/Glass.aiff"
    }]
  }
}

MCP & Custom Commands

# .claude/commands/handoff.md

Summarize this session:
- What was accomplished
- What's left to do
- Key decisions made
- Files changed

Format as a handoff note for
a fresh session to continue.

The Practical Workflow

Tips That Actually Matter

Choosing the Right Model

Different models for different jobs

Pro tip: Use one model to write and a different model to review. Fresh perspective catches bugs the author misses.

Key Takeaways

• Context engineering is the skill — understand token flow, not just context size
• Domain expertise is the multiplier — better knowledge = better prompts = better output
• Use sub-agents for context hygiene — Explore keeps your main session clean
• Start fresh often — long sessions degrade. Use /handoff + new session.
• The model's judgment is usually right — it picks the right tools without micromanagement