Journal entries.

Ryzen 9 9900X, RTX 5070, 64GB. Then a forced power-off cascaded into a corrupted Microsoft profile, OneDrive-induced temp-profile hell, and a from-scratch local account. Closed out by an RDR2 out-of-memory wall. ~3 days from "I have a new PC" to "I have a working PC plus a list."

Six weeks of tool upgrades. Max 5x, Claude Code in the terminal, then Codex joined as the audit layer. Today: cleared Anthropic's Subagents and Agent Skills certificates. Finally have proper names for what I've been doing on instinct. Mental model: a team I can split on demand.

I took everything I'd written from roughly 2004 to 2016, plus three new pieces, and fed them to another Claude instance for corpus analysis. Thirty-odd pieces spanning twelve years. What she pulled out was more precise than I expected — author comparisons, six core signatures, structural weaknesses, a 22-set AB comparison with AI, a training plan. The ten years weren't a blank. They were a different training ground.

Starting point was Jimpj's NTR bot. On the surface boilerplate slop, but the definition layer buries a complete psychological thriller: Beth is the actual perpetrator, Trina is the manipulated subject, the NTR framework is Beth's perception management system. Broke down Beth's four-step manipulation, Trina's shame sequence vs Beth's narrative control. LLM comparison at the end: GLM 5 vs KIMI K2.5 vs Minimax. An ablation study in disguise.