// All posts
Reusing a past compliance answer is an obvious win and a quiet hazard. Here is how I built an RFI reuse engine that reuses safely, with a temporal staleness guardrail that runs on an independent model and fails closed.
I fine-tuned a small model on my compliance notes and it fabricated more, not less. What worked was the opposite of training: a library, citations, and a model that says when the answer is not there.
Anthropic cut 80% of Claude Code's system prompt for the Claude 5 models with no measured loss. I went looking for what that meant for my thirteen agents, and found something worse than bloat: a memory system quietly eating the wrong half of itself.
Three days from scaffold to running system. Thirteen specialist agents, a filing-gate compile pipeline, weekly publishing cron, Korean i18n toggle, agents that ship their own PRs with approval. Plus the discipline not to start V3.
I've shipped a sequence of personal AI projects: Neural, Synapse, Argus. The thinking compounded; the substrate didn't. Here's the bridge.