The Six-Component Harness: A Template for Building Reliably with AI Agents

This is Part 2. Part 1 covers what I learned building Skilldeck with Claude Code — three failure modes, the regression problem, and why instructions aren't enough. This piece is the framework. You can read it standalone, but the story gives it context. Most harness engineering discussions stop at four components: a system prompt, a task list, a progress file, and some tests. That's enough to get an agent building things. It isn't enough to keep a real project coherent across weeks of autonomous sessions, evolving requirements, and features that share code with other features. The harness I ended up with has six components. Each one addresses a specific failure class. The first four are table stakes — the field has largely converged on them. The last two are what make the difference between a harness that works in theory and one that survives contact with a real, evolving codebase. Here's the full template. Component 1: Ground truth File: feature_list.json Failure prevented: Premature c