Uses

The core of the setup. I work solo by directing teams of AI coding agents in parallel — everything else serves this part.

• Claude CodeMy primary agent. It plans, writes, refactors, and carries most of the day-to-day work across every project.
• Codex CLIThe second opinion. A different model reviews the same work adversarially — it catches what the first one rationalizes.
• PlaywrightBrowser QA, driven by agents: real clicks on real pages, with screenshots and console logs as evidence.
• MCP serversThe agents' hands — deploys, browser control, and infrastructure exposed as tools they can call directly.
• SkillsA library of reusable skills — orchestration discipline, debate protocols, digest pipelines — that any agent can load on demand.
• WorkspaceA multi-agent workspace that runs several CLI agents side by side — one screen, several workers, one director.

• AstroStatic-first sites, this one included. Fast by default, and no client-side JavaScript unless a page earns it.
• TypeScriptEverywhere. Types are the contract that keeps a fleet of agents honest inside the same codebase.
• VercelPush to master and production deploys itself. ISR keeps server-rendered signals fresh with no babysitting.
• GitHubThe public identity. The authority of everything here is commits and time — GitHub is the audit trail.

One person, the throughput of a team. The tools only work because the discipline does.

• Prompts as specsA task prompt reads like a spec: scope, constraints, authorization, and a definition of done.
• Cross-model reviewNothing merges on a single model's opinion. A different model reviews adversarially before I do.
• Agent-run QAAgents click through the product in a real browser before anything ships. Screenshots are the receipts.
• Privacy guardsGit hooks enforce an identity contract on every commit — the repo stays clean by machine, not by memory.

• One machineAn Apple Silicon Mac with too many terminal panes open.