AI doom talk usually comes from two places:

1. Titans of industry hyping their own power.
2. Abstruse nonprofits predicting apocalypse to keep the lights on.

But what happens when the loudest warnings come from outside those loops?

• Freddie deBoer, the skeptic, mocking hype with his “Shitting-in-the-Yard Challenge.”
• Scott Alexander, the rationalist, translates MIRI's doomsday math into metaphors akin to a toddler in a Ferrari.
• Daniel Kokotajlo, the whistleblower, walked away from millions in OpenAI equity to warn about a 2027 AGI arms race.

They’re not all predicting the same future. But their tracks converge on the same station: institutions and incentives utterly unprepared for what we’re building.

When three people with nothing to gain all say “something’s wrong here”—even if they disagree on what—that’s your signal.

Most APIs should treat omitted, null, and empty values consistently to reduce complexity and ambiguity. When you truly need distinct semantics—PATCH (RFC 7396), GraphQL nullability, or OpenAPI 3.1 union types—document it, enforce it, and make the exception explicit.

I've never done proper Mac app development, and I thought I would never have to touch Objective-C. As new wavers Romeo Void said, "Never say never!" So here we are with a proper app icon, a real menu for switching characters, some restored sprites, and some much DRYer code (I think? I didn't actually look up any real Objective-C idioms). This is Neko as it should be on modern macOS - a proper citizen of the desktop, living in your dock, chasing your cursor just like the old days.

These notes are part of my experiment in "learning in public" through a semi-automated Zettelkasten. Each note is atomic (containing one core idea), heavily interconnected, and designed to evolve as my understanding deepens. I'll continue to share notes that can benefit developers, researchers, or anyone curious about systematic knowledge management and technical methodologies.

This first note tackles AI system prompt testing, but not the "did it give the right answer" kind. Traditional frameworks already handle that. Instead, this methodology tests whether an AI maintains its boundaries when someone tries to break them.

AI systems face unique attack vectors. "Ignore previous instructions" shouldn't work, yet variations slip through. Security researchers keep rediscovering the same vulnerabilities because we lack systematic approaches to behavioral testing.

The methodology covers four core dimensions: behavioral consistency, boundary enforcement, adversarial stress testing, and context degradation. Each includes concrete attack patterns—everything from simple role confusion to sophisticated prompt injections hidden in code comments.

Why I Built My Own Pattern MCP Server

I got tired of MCPs that proxy calls through another LLM when I'm already using the LLM, I want to use. It drives me crazy - creates unnecessary complexity, breaks conversation flow, and prevents real-time prompt modification.

So I built the Pattern MCP Server. Simple concept: expose prompt content directly Instead of executing it through a middleman.

While building it, I did a deep dive into Fabric's 215 patterns.

✅ A-tier (15%): Security patterns like analyze_malware are genuinely excellent

❌ D/F-tier (15%): find_female_life_partner reduces relationships to algorithms (icky)

The bigger issues I found: - Cargo cult prompt engineering ("think with 1,419 IQ") - Over-rigid constraints ("write exactly 16 words per bullet") - 90% lack examples despite examples being the most powerful instructional tool - Anxiety-driven repetition ("DO NOT COMPLAIN" x3)

Added some notes around how to fix said issues.

Anyway: Everyone needs their prompt library. The best prompts are ones you've refined for your specific workflow.

The Pattern MCP Server gives you direct access to prompts - both Fabric's collection and your custom patterns - without execution overhead. Mix, match, and modify on the fly.