How to Grow by Compounding While Working with AI

This covers a methodology for turning collaboration with AI into more than simple question-and-answer, transforming it into a compounding structure where the results of work accumulate as context for the next session.

Building Context Infrastructure

• Organize all code and knowledge into a systematic directory structure (~/src, ~/vault) so the model can easily search with grep or glob.
• Use MCP (Model Context Protocol) to connect organizational data such as Slack and Drive.
• Write per-project INDEX.md and CLAUDE.md files so the model can immediately grasp the project context, like a new employee.

Encoding Preferences as Settings

• Use CLAUDE.md as a behavioral contract to define the model's tone, way of pushing back, scope of work, and more.
• Manage the scope of settings by dividing them into global, repository root, and project directory levels.
• When files grow long, reduce context costs through lazy loading, specifying only the guide file path.

Automation through Skills

• Turn tasks that repeat at least once a week into workflows by creating markdown files (SKILL.md) that contain the name, trigger, and procedure.
• After performing a task, ask the model to turn that process into a skill, continuously updating it.

A Verification System for Autonomy

• Design verification as a ladder structure, placing low-cost linters at the bottom and high-cost LLM reviews at the top.
• Build self-verification loops where the model runs tests itself or checks browser rendering, among other things.