He keeps a 'boring' single-threaded setup: code editor plus one AI chat, build one feature, release, move to the next, instead of background agent teams.
Per task: scout agent (Opus, low effort, read-only) explores; the top model plans (high effort) then implements (medium) only files in a 'Files touched' list and writes an audit; Opus reviews the diff in a fresh context with Blocking/Non-blocking/Verdict output. Agent files live in ~/.claude/agents/, orchestrated by CLAUDE.md.
When it applies: Claude Code users hitting weekly usage limits
Hetzner VPS, secured with Tailscale and a firewall, SSH via Termius, Claude Code installed, then log in from the phone to ship from anywhere (inspired by levelsio).
When an AI-written spec is hard to picture, ask it for an image-generation prompt of the feature's UI, render it (he used Gemini's image model), keep the parts you like; discard specs you dislike and re-prompt rather than edit them.
In Claude Code use plan mode (Shift+Tab) with Opus as the strategist, then switch to Sonnet to execute the plan step by step, avoiding chaotic one-shot coding.
Maintain files for business context, user context, architectural decisions, constraints and known failure modes; when a long coding session starts contradicting earlier decisions ('context rot'), point the agent at the files instead of re-explaining.
When it applies: building with AI coding agents on long sessions
Because an agent can 'teach to the test' if it sees unit tests, store end-to-end user-behaviour specs separately and only use them to judge whether the software actually works for a real user.
When it applies: AI-native development at level 3+ (agent writes most code)
His design workflow: Paper MCP (Figma for agents) to design, Refero for real-world references so it doesn't look AI-generated, ui.sh agent skills for polish.