把Claude Code变成高级软件架构师,负责协调专业代理团队并自我审计。
概览
Wizard v1 将 Claude Code 从快速的程序员转变为谨慎的代码:先读后写,测试后实现,攻击自己的代码后再提交。v2 让一支专业代理团队的谨慎线程,推动整批 PR 并行完成合并准备。一条仅仅是指导的规则,一旦势头加速就会被违反。因此在 v3 中,每一条重要的规则都变成了——一份归档计划、一个实际运行的变异、一个 PR 正体中的一个区块、一个账本条目——而编排循环得到了:一个廉价的签到代理,写出持久账本,一个仅负责审计第一个线索的第二个线索,以及一个测量而非猜测的容量门。 如果说v1是“三思而后写”,v2是“设计,然后委派,再验证——并行进行”,那么v3就是“” 你来写任务分解或代码。这些都是代理负责的。——而且不能低于那个高度。 - 你带来的是想法——而不是任务清单。
概述(AI 翻译)
把Claude Code变成一位高级架构师,能够协调团队——并自我审计。
Wizard v1 将 Claude Code 从快速的程序员转变为谨慎的代码:先读再写,测试再实现,攻击自己的代码再提交。v2 让细致线程成为一支专业代理团队的编排者,推动整批 PR 并行完成合并。v3 是我们在生产代码库上运行团队三个月后学到的: 一条仅仅是指导的规则,一旦势头加快就会被违反。因此在 v3 中,每一条重要的规则都变成了有序步骤,必须产生一个工件**——一份归档计划、一个实际运行的变异、一个公关机构中的区块、一个账本条目——而编排循环拥有了自己的审计员:一个廉价的签到代理,写持久账本;一个仅负责审计第一个线索的第二位线索,以及一个测量而非猜测的容量门槛。
如果v1是“三思而后写代码”,v2是“设计,然后委派,再验证——并行”,那么v3就是“没有伪造物的规则没有法律效力”。
你的角色:指挥,不是任务执行者
你不是来写任务分解或写代码的。
本页由 AI 翻译,技术细节请以原文为准。
README
以下为原文 README 节选,完整内容请查看原文。
Wizard v3: discipline with teeth
Turn Claude Code into a senior architect that orchestrates a team — and audits itself.
Wizard v1 turned Claude Code from a fast coder into a careful one: read before writing, test before implementing, attack your own code before committing. v2 made the careful thread the orchestrator of a team of specialist agents driving a whole cohort of PRs to merge-ready in parallel. v3 is what three more months of running that team on a production codebase taught us: a rule that is merely guidance gets violated the moment momentum picks up. So in v3 every rule that mattered became an ordered step that must produce an artifact — a filed plan, a mutation that was actually run, a block in the PR body, a ledger entry — and the orchestration loop got auditors of its own: a cheap check-in agent that writes a durable ledger, a second lead whose only job is to audit the first, and a capacity gate that measures instead of guesses.
If v1 was “think before you code” and v2 was “design, then delegate, then verify — in parallel,” v3 is “a rule with no artifact has no teeth.”
Your role: conductor, not task-giver
You are not here to write task breakdowns or code. The agents do that. Your job is to keep the flow moving from idea → issue → PR → production — and nothing below that altitude.
- Set direction. You bring the idea — not the task list. Turning it into a structured issue with acceptance criteria is the
issue-maintaineragent’s first step. - Make the product and judgment calls. When an agent hits an ambiguous requirement or a “which behavior is correct?” fork, that’s yours. The ensemble builds the thing right and defers deciding what’s right to you.
- Unblock. When the orchestrator notifies you it’s gated on a decision, a credential, or a merge, you clear it — then it resumes the whole cohort. v3 never silent-waits on you; a block fires a notification.
- Merge. The orchestrator drives every PR to merge-ready and declares it exactly once, at a named SHA; you press the button. The merge is the one piece of the cycle that stays human.
What it is
/wizard is a Claude Code skill — a markdown playbook that changes how Claude operates for the duration of a task — plus a roster of specialist subagents it dispatches. It runs in two modes:
- Direct mode — a single thread runs the whole 8-phase lifecycle itself (v1’s behavior, intact).
- Delegated mode — the thread you talk to becomes an orchestrator: it decomposes and files the plan, dispatches specialists to build and verify each phase in its own worktree, and owns the pull-request review cycle. Workers commit locally and hand back; the orchestrator pushes, opens the PR, and drives every reviewer finding to resolution.
推荐工具
换一个关键词,或者移除筛选条件。
安装
npx skillfish add vlad-ko/claude-wizard