Wizard v2: the multi-agent workflow

From one disciplined developer to an orchestrated team — without giving up the discipline.

Wizard v1 turned Claude Code from a fast coder into a careful one: it read before writing, tested before implementing, and attacked its own code before committing. That's still the foundation. v2 adds a second gear. When the work is big enough, the careful thread stops being the builder and becomes the orchestrator of a team of specialist agents — an issue maintainer who turns ideas into well-formed issues, an architect who designs, builders who implement in parallel, adversarial critics who try to break it, an independent reviewer that catches what the builders missed, and a review gate that won't let anything merge with an unanswered finding. One thread becomes a pipeline that drives a whole cohort of work to merge-ready at once.

If v1 was "think before you code," v2 is "design, then delegate, then verify — in parallel, without idling." And the most important shift is for you: you stop being a task-giver and become a conductor.

Your role: conductor, not task-giver

This is the defining change of v2. You are not here to write detailed task breakdowns, spell out implementation steps, or write code. The agents do that. Your job is to keep the flow moving from idea → issue → PR → production — and nothing below that altitude.

Concretely, the human conductor:

• Sets direction. You bring the idea ("we need transfer-status tracking") — not the task list. Turning that idea into a structured issue with acceptance criteria is the issue-maintainer agent's first step, not your chore.
• Makes the product and judgment calls. When an agent hits an ambiguous requirement, a tradeoff, or a "which behavior is correct?" fork, that's yours to answer. The ensemble is deliberately good at building the thing right and deliberately defers deciding what's right to you.
• Unblocks. When the orchestrator pings that it's gated on a decision, a credential, or an external dependency, you clear it — then it resumes the whole cohort.
• Merges. The orchestrator drives every PR to merge-ready and declares it exactly once; you do the final merge. v2 never auto-merges. The merge button is the one piece of the cycle that stays human.

Everything between "idea" and "merge button" — issue authoring, design, building, reviewing, answering review findings — is delegated. You conduct; the ensemble plays.

What it is

/wizard is a Claude Code skill — a markdown playbook that changes how Claude operates for the duration of a task. v2 runs in two modes:

• Direct mode — a single thread runs the whole 8-phase lifecycle itself (this is v1's behavior, kept intact).
• Delegated mode — the thread you're talking to becomes an orchestrator: it designs the change, dispatches specialist subagents to build and verify it, and owns the pull-request review cycle. The boundary between worker and orchestrator is the git commit — workers commit locally and hand back; the orchestrator pushes, opens the PR, and drives every reviewer finding to resolution.

The complexity of the work decides which mode you get. A one-line fix never pays the multi-agent tax; a multi-domain feature with shared state and lifecycle transitions gets the full ensemble.

The ingredients

/wizard isn't just a prompt — it's a workflow built on specific ingredients that work together.

Carried over from v1 (still the foundation):

1. CLAUDE.md — your project's rules file. Coding standards, naming conventions, architecture decisions, anything Claude should always know. /wizard reads this first, every time. It's also the only thing a dispatched subagent inherits besides its brief — so it's load-bearing.

2. Issues as the source of truth — every feature or bug gets an issue (or epic) before coding starts, with acceptance criteria. /wizard tracks progress by checking off boxes at merge-time, and references the issue in every commit. In v2 this is the first agent step, not a human chore: a dedicated issue-maintainer agent turns your raw idea into a consistent, well-structured issue or epic — uniform labels, real acceptance criteria, and native parent↔sub-issue links with a live progress bar. The cycle starts at idea → issue (by the maintainer), so you never hand-write a ticket.

3. Codebase-first exploration — before writing a line, /wizard reads the existing code, greps for methods and relationships, and verifies assumptions. No hallucinated APIs.

4. TDD, no exceptions — failing tests first, then minimal implementation, then verify. With a mutation-testing mindset: assertions that break if the code changes, not just assert(worked).

5. Branch-per-concern — clean branch, focused PR, one concern at a time. No stacked branches, no tangled dependencies.

New in v2 (the orchestration layer):

1. The orchestrator/worker split — for delegated work, responsibility splits at the git commit. The worker builds and commits locally; the orchestrator verifies the diff, pushes, opens the PR, and monitors. This two-phase boundary is what lets the orchestrator catch a bad diff before it's exposed, compose the PR with cross-cut context, recover cleanly from a crashed worker, and be the single owner that declares merge-ready exactly once.

2. The agent ensemble — gate-routed and mediated. A complexity gate fires first and classifies the work; trivial work takes a single subagent, complex work gets the full chain: persona critic lenses + a doc librarian harden the requirements, an architect designs the change and writes the failing-test spec (but no production code), backend and frontend specialists turn the spec green in parallel, and a QA engineer plus the lenses verify the result. The agents that build are never the ones that sign off — generator ≠ evaluator.

3. The parallel pipeline (a cohort of up to ten) — the orchestrator doesn't drive one PR at a time. It selects a cohort of up to ~ten issues and drives them ALL to merge-ready concurrently — each in its own isolated worktree with its own PR — and refills the cohort as PRs merge. Wait windows (a reviewer pass, the CI suite, a quiescence window) are work-time: while one PR waits, the orchestrator advances the others — sweeping worktrees, auditing findings, and spawning the next candidate up to the per-author depth band. Idle is forbidden while independent work exists.

4. The AI-review gate (non-negotiable, and the loop closes) — every PR is reviewed by an independent, dedicated AI reviewer that did NOT build the change (CodeRabbit or similar). Because it's a fresh set of eyes with no stake in the implementation, it reliably catches things the internal team missed. After opening the PR, /wizard reads every finding across all three surfaces (inline, review body, issue-level summary) and routes each one back to the team: a valid finding becomes a real fix dispatched to the specialist whose layer it lives in (backend → backend-expert, view → frontend-expert, and so on); a false positive gets a reply-and-resolve, separating the premise (often right) from the suggested remedy (often wrong for your actual stack). The loop runs until the status is clean and a quiescence window has elapsed. No unresolved findings, ever. That closing of the loop — every finding fed back and answered — is what makes the ensemble a team rather than a pile of agents. Then it declares merge-ready and hands the merge to you.

Each phase has a checkpoint. Claude won't rush ahead.

The difference

Without /wizard:

You: "Add a transfer-status tracking feature"

Claude: writes 400 lines, misses a race condition, hard-codes a string that should be a constant, skips tests

With /wizard (delegated mode, a complex feature):

You (conductor): "we need transfer-status tracking" — just the idea

Claude (orchestrator): dispatches the issue-maintainer to turn the idea into a structured issue with acceptance criteria

You: /wizard implement the transfer-status feature in the issue

Claude (orchestrator): runs the complexity gate → dispatches persona lenses + a doc librarian to harden the ACs → dispatches the architect to design it and write the failing-test spec → fans the build out to backend ∥ frontend ∥ QA in parallel → re-runs the lenses adversarially against the built diff → pushes, opens the PR → an independent AI reviewer reviews it → routes each finding back to the team (real fix to the owning specialist, or reply+resolve for a false positive) → waits out quiescence → declares merge-ready and pings you to merge — all while driving the rest of the cohort in parallel

You: merge

The output is the same — working code. But it ships without the 2am "why is this broken in production" follow-up, and a whole cohort of PRs ships at once instead of one at a time — while you stay at the conductor's altitude.

Upgrading from v1

v2 is a superset. If you used v1, the 8-phase methodology and its TDD/adversarial-review core are all still here — that's "direct mode." v2 adds the orchestration layer on top, and it only engages when the work is complex enough to warrant it. You don't have to change how you invoke /wizard.

What's new in your install: an agents/ directory of specialist role definitions, a reference/ directory of deep-dive docs (the threading model, the parallel pipeline, the PR review cycle), and an ARCHITECTURE.md with the system diagrams. The installer drops the agents into your .claude/agents/ so the orchestrator can dispatch them.

Still want v1? It's preserved at the v1 git tag:

git clone https://github.com/vlad-ko/claude-wizard
cd claude-wizard && git checkout v1

Or install the v1 skill files directly from that tag's raw URLs.

Install

One command from your project root:

curl -sL https://raw.githubusercontent.com/vlad-ko/claude-wizard/main/install.sh | bash

This installs the /wizard skill into .claude/skills/wizard/ and the agent roster into .claude/agents/.

Or manually:

# Skill
mkdir -p .claude/skills/wizard
for f in SKILL.md CHECKLISTS.md PATTERNS.md; do
  curl -sL "https://raw.githubusercontent.com/vlad-ko/claude-wizard/main/skill/$f" -o ".claude/skills/wizard/$f"
done

# Reference docs (loaded on demand by the skill)
mkdir -p .claude/skills/wizard/reference
for f in threading-model.md parallel-pipeline.md pr-review-cycle.md; do
  curl -sL "https://raw.githubusercontent.com/vlad-ko/claude-wizard/main/reference/$f" -o ".claude/skills/wizard/reference/$f"
done

# Agent roster
mkdir -p .claude/agents
for f in architect backend-expert frontend-expert qa-engineer doc-librarian issue-maintainer domain-user-lens; do
  curl -sL "https://raw.githubusercontent.com/vlad-ko/claude-wizard/main/agents/$f.md" -o ".claude/agents/$f.md"
done

Usage

In Claude Code, type:

/wizard implement the user authentication flow described in the issue

Claude responds with ## [WIZARD MODE] and begins the phased approach, signaling each transition:

## [WIZARD MODE] Phase 1: Understanding & Planning
...
## [WIZARD MODE] Phase 3: Test-Driven Development
...

You can also invoke it mid-conversation:

/wizard this is getting complex — let's be more systematic about this

Customizing the agents

The agent roster is generic by design. Two files need your attention before the ensemble fits your product:

• agents/domain-user-lens.md is a TEMPLATE, not a ready agent. It embodies one user persona's adversarial point of view — and your product's personas aren't anyone else's. Copy it once per distinct persona in your product (e.g. admin-lens.md, end-user-lens.md, power-user-lens.md), and fill in each persona's real surfaces, domain rules, and failure modes. The template ships with three neutral example personas to show the shape.

• agents/backend-expert.md and agents/frontend-expert.md point at "your project's CLAUDE.md" for the framework-specific rules. They're written to reference your standards rather than embed a particular stack — so the more complete your CLAUDE.md, the sharper they get.

Everything references "your test runner / your CI / your review bot" rather than a specific stack. The methodology is the product; the stack is yours.

What's included

How it works

/wizard wires the ingredients above into an enforced sequence — read the rules, define "done," explore the code, write failing tests, implement minimally, verify, self-review adversarially, then open a PR and drive every finding to resolution. In delegated mode it adds a gate-routed, orchestrator-mediated agent ensemble on top, with a parallel pipeline that keeps wait windows productive.

There's no magic. It encodes the habits of senior engineers into a repeatable process. Claude doesn't lack the ability to do these things — it lacks the process to do them consistently. /wizard is that process.

Contributing

This project is small, opinionated, and hungry for fresh ideas. PRs welcome and encouraged.

Ways to contribute:

• Framework overlays — add a frameworks/<stack>/ directory with stack-specific phase additions people can merge into their SKILL.md.
• Persona lenses — contribute a well-written domain-user-lens instance for a common product shape (a SaaS admin, an e-commerce shopper, an API consumer).
• New patterns — found a bug pattern /wizard should catch? Add it to PATTERNS.md.
• Phase improvements — battle-tested a refinement? Open a PR with a before/after example.
• Bug reports — if /wizard missed something it should have caught, that's a bug in the prompt. File an issue with the scenario.

How to contribute: fork → change → open a PR with what changed and why. Bonus points if you use /wizard to make the PR.

Origin

This workflow was distilled from months of production use orchestrating Claude Code on a real codebase — hundreds of PRs, real race conditions caught by the adversarial-review phase, and review-bot findings that would have reached production without the quality gate. The stack-specific machinery has been stripped out; the architecture and the why are what's captured here. The methodology works with any language, framework, or stack.

Requirements

• Claude Code CLI (the ensemble uses its subagent dispatch primitives).
• A git repository (the skill uses your host's CLI for issue/PR integration).
• An automated code-review bot for the review gate (CodeRabbit or similar). The cycle works without one, but the gate is where /wizard really shines.

License

MIT. Fork it, adapt it, make it yours.