What I've built
Short clips of personal AI agent tooling: real features, no mockups.
I'm Fernando, a software engineer. In my own time I run several Claude Code sessions at once on one Linux machine that stays on, and I have built a small system around them. A vault holds my notes and projects. Around three hundred skills, commands, hooks and tools are the procedures an agent can run.
A “drive” carries a project through its stages while I sleep, and when it needs me it opens a decision page. A cockpit shows every session and which one is waiting for me. The skills graph is the map of all of it, generated from the skills' own files. Each clip below is one part of that setup, filmed while it ran.
Two agents, one project, no collision
A coding agent noticed another already working on my project, split the work with it in one message, and only then asked me to approve.
Built 2026-09-28 · Filmed 2026-09-28
Watch the clip →
Agent questions as annotated web pages
When an agent needs my sign-off it builds a page: a hover glossary for the jargon, and comments anchored to the exact sentence I am reacting to.
Built 2026-08-06 · Filmed 2026-10-14
Watch the clip →
A searchable map of my agent skills
Nodes and links are generated from the skills' own metadata, so the map never goes stale. Search a skill and only its real connections remain.
Built 2026-06-11 · Filmed 2026-09-26
Watch the clip →
I dictated a review, and the page was rebuilt from it
A long voice review replayed from what the form saved, then a changelog page rebuilt from my notes: about 10 MB down to about 3 MB.
Built 2026-09-28 · Filmed 2026-09-29
Watch the clip →
One slice of the map, four layouts, same data
A nine-node slice of the skills graph is re-laid out four ways, live, so exploring never distorts what is actually connected.
Built 2026-07-15 · Filmed 2026-10-16
Watch the clip →