Coding agents are fast, but someone still has to carry context from spec to code to review. The workbench moves every feature through a fixed sequence of phases (spec, requirements, acceptance criteria, build, audit, verification, merge) and brings humans in only at defined gates.

Before any code is written, Claude's frontend-design skill turns the approved spec into a UI/UX prototype of every new screen and state. Once I approve it, the prototype becomes part of the plan. After the build, a design-fidelity audit has a vision-capable model compare the implemented screens, captured with Playwright, against the prototype and the spec. Layout, copy, design tokens and states are checked, and high-severity deviations block the merge.

Each phase runs in its own Claude session with role-specific tool permissions. Features build in parallel in isolated git worktrees, and independent security and quality scans run bounded fix loops. A hub, workers and a desktop app track cost against per-feature and daily budgets.

Schema-enforced structured outputs, offline evals graded against a hand-labelled corpus, live canary projects and property-based tests keep the agents honest. More than 2,200 commits in its first seven months. The screens shown here preview the next, kanban-style interface.