This entire blog series was drafted by Claude inside a sandboxed pod; the sandbox got rebuilt three times — Docker, mitm plus credential proxy, then a K8s pod — to make that safe.
How I wired up a Claude agent that reads engineering task-logs and produces blog post drafts and GraphKnow wiki sources — then used it to write this post.
I shipped a polished document-RAG chat clone, never opened it, and pivoted to a personal LLM-maintained wiki — same data plane (Mongo / MinIO / Qdrant / Neo4j), different surface, four new entities, Slice 0 deployed.
An AlphaZero-style pipeline for the augmented chess engine ran into 512Mi pod limits, and the fix taught me the difference between liveness and readiness probes.
Six weeks of multiplayer bugs across three games converged on the same contract — deterministic appliers, atomic sequence allocation, server-owned identity.
