Claude Code gets a multi-agent team with parallel PRs and TDD

V.j.k.'s original `/wizard` workflow addressed what they frame as AI coding assistants' core weakness: not intelligence, but process. V1 forced a single Claude instance into disciplined senior-developer habits — reading the existing codebase before writing a line, defining acceptance criteria upfront, writing a failing test first, implementing the minimum code to pass it, then attempting to break the result. This eliminated the "why is this broken in production" class of problems, but it was still a single-threaded individual contributor with a ceiling of one task at a time.

V2 reframes the mental model entirely: instead of a senior developer, the system becomes a senior architect running a team.

V2 reframes the mental model entirely: instead of a senior developer, the system becomes a senior architect running a team. A main-thread orchestrator — which never writes code itself — fans work out to specialist subagents for backend, frontend, and QA, each working in its own isolated git worktree and building in parallel. An independent reviewer subagent (one that did not build the code) handles the review gate, routing findings back to the specialist responsible for each layer across all open PRs simultaneously. A separate issue-maintainer agent converts a one-line idea into a fully structured ticket with title, acceptance criteria, labels, and epic linkage, removing the manual ticket-writing step that the author identifies as a quiet bottleneck. The system supports up to ten pull requests open and moving concurrently.

V1 is preserved verbatim inside v2 as "direct mode," used for simple one-line fixes where spinning up the full team would not be worth the overhead. The author notes that their own role across the two versions shifted in two stages: in v1 they stopped writing code and doing line-by-line reviews; in v2 they also stopped writing deep technical prompts and GitHub issues, leaving them to tune workflows and supervise the agents — a role they describe as "conductor." The source text is truncated before the full technical walkthrough of each agent component.

Key facts

1. 01The original `/wizard` prompt (v1) enforced a read-first, TDD, self-review loop on a single Claude Code instance.
2. 02V2 introduces a main-thread orchestrator that never writes code itself, instead dispatching to specialist subagents.
3. 03Subagents each operate in their own isolated git worktree and build in parallel.
4. 04The system can run up to ten pull requests open and moving concurrently.
5. 05An issue-maintainer agent converts a one-line idea into a structured ticket with acceptance criteria, replacing manual prompt-writing.
6. 06An independent reviewer subagent — one that did not build the code — handles the review gate and routes findings back to the responsible specialist.
7. 07V1 is preserved as 'direct mode' inside v2 for simple tasks where the full team overhead is not warranted.

Topics