My Toxic Relationship with AI: The Pitfall of Perfect Context Engineering

I have to yell at my AI to get it to use the right context every time.

I added 15 “MANDATORY” labels in all caps to my skills. I ran the same prompt 10 times just to see if my skill would trigger consistently. And every time it failed, I pushed a commit to “fix it.”

I didn’t want a useful AI. I wanted an infallible AI.

I’m that person in a relationship who demands absolute perfection.

A mistake? Unforgivable. I want determinism.

I’m in a toxic relationship with AI.

And this toxicity has cost me time, energy… and way too many commits.

In this talk, I’ll share the story of this relationship in three acts:

• The honeymoon (10 min) — The early adopter’s excitement and the promises of context engineering
• The control freak (15 min) — The obsession with perfect context, the 15 “MANDATORY” labels in all caps
• Couples therapy (15 min) — The breakthrough: you can optimize your context in healthier ways. And AI isn’t a deterministic remote control—it’s an iteration partner that already makes me 3x more productive

I’ll show you excerpts from toxic skills, the successive diffs from my obsessive PR, and the healthy, efficient, maintainable version.

By the end of this talk, you’ll walk away with:

• Examples of context files that work — and ones that sabotage
• Warning signs of over-engineering — when your “MANDATORY” labels become counterproductive
• Best practices beyond context files — choosing the right interaction mode, managing the context window
• A realistic mindset — even with occasional misses, a well-calibrated context is still a valuable asset for building clean, shippable code faster