Vexjoy

n. The specific satisfaction found in solving deeply frustrating technical problems.

Adaptive Thinking Doesn't Change Quality. It Changes Variance.

I blind-tested Claude Code’s CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING setting with 20 headless sessions. Quality was identical. Variance was not.

April 7, 2026 · 6 min

Claude Code Memory: Progressive Discovery vs. Context Dumps

Most AI memory systems dump everything into context. Ours uses progressive discovery: a nightly dream cycle distills all learnings into ~2000 tokens.

April 6, 2026 · 8 min

Tell the Model What to Do, Not What to Avoid

Blind A/B tests on Go code generation showed positive framing in agent instructions consistently outperforms prohibition language. I built a linter to enforce it.

March 28, 2026 · 4 min

I Was Excited to See Someone Else Build a /do Router

Nine months of building in public with limited traction. Then someone posted a system with the same architecture and I felt something I wasn’t expecting.

March 22, 2026 · 7 min

The Handyman Principle: Why Your AI Forgets Everything

Context is a scarce resource, not a dumpster. Here’s how I stopped Claude from forgetting.

January 2, 2026 · 5 min

Everything That Can Be Deterministic, Should Be: My Claude Code Setup

The alien tool came with no manual. Here’s how I learned to hold it.

December 29, 2025 · 7 min

The /do Router: Keyword Matching for Specialist Selection

A 394-line routing system that loads domain specialists instead of letting Claude improvise. 33 agents, 57 skills, keyword matching. The mechanism and failure modes.

December 24, 2024 · 8 min