Most of what’s on this site — including this write-up, and the design system the site itself is built on — was planned, implemented, and verified by an AI-native development environment I built across the project workspace to plan, implement, and verify changes.
What’s actually running
A tmux “mission control” layer brings up pre-configured terminal sessions and monitors across every project with one script, so a fresh terminal lands in a known-good layout instead of a blank shell.
Two always-on agents sit on top of that with distinct jobs. One is a lightweight notifier/gateway — scheduled jobs, quick status queries, nothing that touches code. The other is a “foreman”: it takes an instruction, spins up an isolated coding session, branches the relevant repo, and opens a pull request for human review. Neither agent merges its own work.
The planning pipeline
The part that actually ships features is a multi-stage pipeline: an AI planner researches a goal and writes it down as a session-by-session checklist, each session sized to fit in one clean AI context; an AI orchestrator runs the checklist, handing each session to a fresh implementer that both writes the change and verifies it before moving on; once every session lands, the orchestrator validates the whole feature end-to-end against the original goal; a wrap-up pass triages whatever the review surfaced into the next round of plans; and a ship pass carries the validated result through its release checks. The loop has shipped tested features across ClaudeD, the private dashboard, and this site — including this site’s design system and content pipeline.
One source of truth
There’s no external task-tracking SaaS anywhere in this. Priorities and in-flight work live as plain markdown kanban files, versioned in git; a private mission-control dashboard and the agents themselves all read and write the exact same files, so there’s never a sync step or a second copy to drift out of date. The dashboard runs as a mobile-friendly private web app, restricted to my own private network, and adds service health, git activity, board views, and agent run history on top of the same file layer. Capturing a new item is a single voice or text note; an AI pass files it onto the right list within seconds instead of me copying it somewhere by hand.
The pattern worth naming
Strip away the specific tools and the shape underneath is: AI plans, AI implements, AI verifies, and a separate AI orchestrator validates the result against the original goal — with me holding the review gates instead of reading every line of every diff. That’s the same loop I use to build ClaudeD and this site.