*This article was created using voice input with AI (Aqua Voice) for text conversion. Please note there may be some inconsistencies.
The Era of Controlling AI Through “Requests” Is Over — Harness Engineering Practical Guide
Hello. How’s everyone doing?
I’m Hiro from Full Body Tracking Lab.
In the previous article, I explained the three types of AI engineering. Today I’m going into the practical implementation of harness engineering — how to actually build one.
What a Harness Looks Like in Practice
A harness consists of several structural components that together create an environment where AI operates reliably.
CLAUDE.md — The Foundation File
This is the master document AI reads on every startup. It contains project structure, rules, workflows, and references to other files. Think of it as the blueprint for your AI’s behavior.
Skills Files
Persona definitions that tell AI who it is and how it should behave. For my writing, this includes: first person is “I,” tone is casual-polite, always include a hidden theme, never recommend devices I haven’t personally tested.
Memory System
Decision logs, session learnings, error records — all persisted as files that survive between sessions. AI reads these on startup and maintains continuity.
Quality Hooks
Automated scripts that run after AI produces output. They check for rule compliance: correct persona, proper format, factual accuracy against the knowledge base. Failures trigger automatic correction loops.
Workflows
Predefined step sequences for common tasks. For article writing: theme confirmation, hidden theme design, structure, drafting, revision. AI follows these steps without needing to be told each time.
The Key Insight: Structure Over Requests
In context engineering, you tell AI “please write in this style.” That’s a request. AI might follow it, or might drift.
In harness engineering, the style is defined in files that AI reads automatically. Checkers verify compliance. If AI drifts, it gets corrected structurally, not through another request.
This is the difference between asking someone to drive carefully versus building guardrails on the road.
Building Your First Harness
Start small:
- Create a CLAUDE.md with your project’s basic rules
- Add one skill file defining your writing persona
- Set up a simple memory file for recording decisions
- Add one quality check (even just verifying first person usage)
Expand from there as needed. The harness grows organically as you identify patterns that need structural enforcement.
Thank you for reading! Please follow us on X and subscribe to our YouTube channel!