Collin Wilkins / AI systems / practical automation
Practitioner notes on AI tools, convention files, and automation architecture.
Lead Software Engineer with 12 years across Ford Motor Company and enterprise financial data. I write about what actually works when engineering teams adopt AI coding tools, and what breaks when they don't set up the context layer first.
Start Here
View all 46 articles →
AI & Agents
Context Engineering: The Skill That Makes AI Coding Tools Actually Work
Why AI tools produce bad output isn't the model, it's the context. Convention files are how you turn AI coding tools from novelty into compounding productivity.
Read article
AI & Agents
LLM Gateway Architecture: When You Need One and How to Get Started
When you need a gateway, when you don't, and how to avoid lock-in. The routing layer most teams skip until it's too late.
Read article
AI & Agents
The Claude Code Productivity Paradox
Individual metrics are up. Organizational metrics are flat. The gap between 'developer goes faster' and 'team ships more' is where most AI pilots die.
Read articleLatest
Browse archive →
AI Code Quality: Bad Code Is an AI Tax
AI makes code easy to generate, but bad architecture creates review drag, retry loops, trust issues, and expensive course correction.
What Is an AI Agent Harness?
The model is the brain. The harness is the operating environment that makes the brain useful. A plain-English breakdown of agent loops, tools, permissions, memory, verification, and why harness design matters more than model choice.
Managing in the AI Era Is Harder Than It Looks
Pull requests are up 20%. Incidents per PR are up 23.5%. The supervision gap is real and most managers are missing it.
Coding Wars 2.0: The Definitive Comparison of Kimi K2.6, GLM-5.1, and Claude Opus 4.7
The gap between Kimi K2.6 and GLM-5.1 on public coding benchmarks is basically noise. The real decision is when to route work to them versus Claude Opus 4.7.
Taste Is a Moat
Taste isn't mystical. It's System 2 thinking, run on the same decision so many times it drops into System 1. That's why AI can't get there — and why it's a moat.