Early decisions from setting up a pleasant launch cockpit where a creative collaborator can run observable experiments without first becoming a project manager.

A WhatsApp screenshot, a live CMS, a repository full of operating context, and four annotated images. This is what context engineering looks like when it helps a real person use new software.

Voice dictation, computer use, terminal agents, parallel work, client systems, prompt governance, private telemetry, and a way to check whether speed is improving or hiding my judgment.

A 37-day trace of how I work with agents: hundreds of one-turn dispatches, a small number of deep systems arcs, and 6.4 billion cache-heavy tokens behind them.

Agent/tool-layer infrastructure: CLI modules with isolated stores and NATS event sourcing, so humans and agents drive the same tools through the same surface. Python, Docker, Kubernetes.

Building a resume-safe, multi-pass system that turns 4,700 unsorted iPhone photos into a categorised library - using metadata heuristics, perceptual hashing, and Claude's vision API.

I pointed a tokeniser at the repo. The boot sequence had an elephant in it.

I asked an agent to write about agents producing competent-sounding-but-wrong output. It produced competent-sounding-but-wrong output.