Articles
Practical write-ups from building and running a persistent AI workflow with Claude Code — agent memory, autonomy, and the failure modes that only show up in daily use.
Isolating Autonomous Agent Jobs from Prompt Injection
Prompt injection in a background agent is a tool-access problem, not a wording problem. The isolation pattern — an inert LLM step, hardcoded destinations, wrapped input, and a human gate — that keeps an unattended job safe, and where it still fails.
Read articlePersistent Memory for AI Agents Without a Vector Database
You don't need RAG to give an agent long-term memory. The session-lineage architecture — plain markdown, a bootloader, and a session-end hook — that gave Claude Code continuity across 250+ sessions, and when to use a vector database instead.
Read articleBuilding a Personal AI Operating System
A field report on running my real work — billing, HR, sales, client demos — through Claude Code every day. What held up, and what broke.
Read case study