Multi-agent orchestration for full-stack monorepo builds. Audits a PRD against a design handoff (HTML, screenshots, claude.ai/design output) to find gaps, writes an execution plan, and dispatches parallel subagents through 4 phases. Coordinates ALL 16 superdev skills — design-preservation (when source is a prototype, not Claude Design), design-to-nextjs (shadcn translation when source is Claude Design), nestjs-enterprise-backend, security-review-and-fix, prototype-to-saas, exploratory-qa, systematic-debugging (on any bug found mid-build), product-completeness-audit (between QA and ship), brutal-exhaustive-audit (final pass before declaring done). Reads `.claude/memory/superdev-learned/` before every subagent dispatch and threads project-specific lessons into agent prompts so the system avoids repeating past mistakes (self-learning loop via superdev-self-learning skill).
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Multi-agent orchestration for full-stack monorepo builds. Audits a PRD against a design handoff (HTML, screenshots, claude.ai/design output) to find gaps, writes an execution plan, and dispatches parallel subagents through 4 phases. Coordinates ALL 16 superdev skills — design-preservation (when source is a prototype, not Claude Design), design-to-nextjs (shadcn translation when source is Claude Design), nestjs-enterprise-backend, security-review-and-fix, prototype-to-saas, exploratory-qa, systematic-debugging (on any bug found mid-build), product-completeness-audit (between QA and ship), brutal-exhaustive-audit (final pass before declaring done). Reads `.claude/memory/superdev-learned/` before every subagent dispatch and threads project-specific lessons into agent prompts so the system avoids repeating past mistakes (self-learning loop via superdev-self-learning skill).
PRD ↔ Design Audit + Build Orchestrator
A multi-agent system for converting a PRD + a design handoff into a shipping full-stack monorepo. The orchestrator (the main Claude in Claude Code) does no implementation work itself — it dispatches specialized subagents and coordinates phases.
🧠 Self-learning — the orchestrator reads .claude/memory/superdev-learned/ BEFORE every subagent dispatch and threads relevant lessons into agent prompts. The system learns from past mistakes in THIS project. See superdev-self-learning.
🎨 Design preservation — when the user's source is a prototype (not Claude Design output), the orchestrator inserts the design-preservation Phase B.0 to copy the source verbatim and gate every Phase C wave on a fidelity audit. Pixel drift > 1% blocks the wave.
Mandatory pre-dispatch checklist (every subagent, every time)
Before dispatching ANY subagent, the orchestrator does:
# 1. Read learned lessonsls .claude/memory/superdev-learned/ 2>/dev/null
# 2. For each lesson, check if applies_to_agents includes the about-to-dispatch agent# 3. Append matched lessons to the agent's prompt under "## Lessons learned in this project"# 4. Announce to user (first dispatch of the session only):# "Found N learned lessons for this project. Threading them into dispatches."
product-completeness-audit (between QA and audit) THEN brutal-exhaustive-audit
Distinguishes demo-vs-product, then dots every i
Capstone of Phase D — "rate each module 1–10 / make it production-ready / complete it fully"
production-readiness-audit (the iterative loop: 5 passes → score every module → fix → re-audit until all ≥ 9)
Turns "looks done" into a scored, proven, ledger-enforced "done"; subsumes completeness+brutal into one scored loop and writes gates.readiness
User-frustration signal detected OR verifier rejection
superdev-self-learning (auto-dispatched by hook)
Capture lesson; future dispatches inherit
When NOT to invoke a skill:
❌ Don't invoke brutal-exhaustive-audit for a single-feature edit (it's whole-product)
❌ Don't invoke design-preservation for Claude Design output (defeats translation purpose)
❌ Don't invoke systematic-debugging for refactors with no bug to chase
❌ Don't invoke product-completeness-audit mid-build (run only between QA pass and ship claim)
❌ Don't invoke frontend-refactoring for multiple modules at once — one module per dispatch, atomic. For N fat modules, run N separate dispatches.
❌ Don't dispatch frontend-rewirer on files > 300 lines — frontend-refactoring must run first (the rewirer itself refuses).
When to use this skill
Use whenever the user has:
A PRD (markdown, docx, pdf, or even pasted text) describing a product
A design handoff (HTML from claude.ai/design, screenshots, a .zip bundle, or even just verbal spec)
The intent to build BOTH a Nest.js backend AND a Next.js frontend in one monorepo
Do NOT use this skill for:
Single-side builds (frontend-only or backend-only) — invoke design-to-nextjs or nestjs-enterprise-backend directly
PRD-without-design (no UI to audit against — write the design first, or skip the audit phase)
Existing codebases adding a feature — use the per-skill recipes; this is for greenfield orchestration
How to invoke this skill
This skill is designed for Claude Code (the CLI), where it has access to the subagent system. The skill installs 10 specialized subagent definitions into .claude/agents/ and the main session orchestrates them.
Three invocation patterns, from least to most opinionated:
Pattern 1 — natural language (default, recommended)
After installing the skill, just describe what you want in a normal Claude Code session:
I have a PRD at docs/PRD.md and a design handoff bundle at design/.
Build the full-stack app.
The main session reads this skill's SKILL.md, runs Phase A.1 to install the subagents into .claude/agents/, and then drives the four phases by delegating to subagents through natural language ("Use the prd-analyst subagent to read the PRD and produce PRD_DIGEST.md") or @-mentions (@"prd-analyst (agent)" read the PRD).
Pattern 2 — main session AS the orchestrator (claude --agent)
If you want the entire session to behave as the orchestrator (its system prompt replaces the default Claude Code prompt), launch with:
claude --agent prd-design-build-orchestrator
This requires a prd-design-build-orchestrator.md agent file at the project or user level. The skill's install step writes one automatically if you opt in (see references/install-orchestrator-as-agent.sh).
Pattern 3 — interactive setup via /agents
In an existing session, run /agents to open the agent management UI, navigate to the Library tab, and confirm all 10 subagents are present. Use this to inspect or edit agent definitions before kicking off Phase A.
Optional: agent-teams mode for specific phases
This skill defaults to subagents (stable, lower-cost, results-only). Some phases benefit from agent teams — independent Claude sessions that can message each other and be steered individually — particularly the audit phase and feature-by-feature builds. Agent teams are experimental and require CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in settings.json or environment.
See the "Agent teams (optional)" section near the end of this skill for which phases can use teams and how to invoke them. The agent definitions in .claude/agents/<name>.md work as both subagent definitions AND teammate types without modification — when the user asks the lead to "spawn a static-auditor teammate," the same definition file applies.
Cross-cutting conventions enforced by every agent
These conventions span both apps and every feature. The orchestrator surfaces violations during wave gates.
shadcn/ui is THE component substrate — no other UI libraries
Every visual primitive in the frontend comes from shadcn/ui via @/components/ui/*. No Headless UI, no raw Radix imports, no MUI, no Mantine, no Chakra, no Ant Design, no DaisyUI, no Bootstrap, no react-bootstrap, no semantic-ui, no flowbite, no nextui, no tremor. Hand-rolled buttons/inputs/dialogs are also forbidden — if shadcn has it, use shadcn's.
What this includes:
Every form input: <Input>, <Textarea>, <Select>, <Checkbox>, <Switch>, <RadioGroup>, <Slider> from @/components/ui/*. The form itself uses <Form> from shadcn (which wraps react-hook-form).
Every button: <Button> with shadcn's variant API (variant="default" | "destructive" | "outline" | "secondary" | "ghost" | "link").
Every overlay: <Dialog>, <Sheet>, <Drawer>, <Popover>, <HoverCard>, <Tooltip>, <AlertDialog> from shadcn.
Every table: shadcn's <Table> from @/components/ui/table (paired with TanStack Table for data plumbing, but the visual primitives are shadcn).
Every menu: <DropdownMenu>, <ContextMenu>, <Menubar>, <NavigationMenu> from shadcn.
Every list/data display: <Card>, <Badge>, <Avatar>, <Skeleton>, <Separator>, <ScrollArea> from shadcn.
Every sidebar uses the shadcn sidebar block (<Sidebar>, <SidebarProvider>, <SidebarTrigger>, <SidebarMenu>, <SidebarMenuItem>, <SidebarHeader>, <SidebarContent>, <SidebarFooter> from @/components/ui/sidebar). No custom drawer/aside layouts.
Every notification: <Sonner> (sonner via shadcn) or shadcn's <Toast>.
Every command palette: shadcn's <Command> / <CommandDialog>.
What's allowed alongside shadcn:
lucide-react icons (shadcn's default icon set)
TanStack Table, TanStack Query (data plumbing, no visual primitives of their own)
react-hook-form (wired through shadcn's <Form>)
zod (validation, no UI)
cmdk (used by shadcn's Command)
sonner (used by shadcn's toast)
recharts for charts (shadcn doesn't ship charts as primitives; the shadcn chart wrapper around recharts is acceptable, otherwise raw recharts)
Tailwind utility classes for layout, spacing, color application
Forbidden import paths (the ui-auditor agent enforces this; see below):
Setup is owned by monorepo-bootstrapper (Phase B): runs pnpm dlx shadcn@latest init, then bulk-adds every primitive plus the sidebar block. By the time frontend-module-builder runs in Phase C, every primitive is already in apps/web/src/components/ui/.
Enforcement is owned by ui-auditor (Phase C wave gates + Phase D): greps for forbidden import paths, raw <button> / <input> / <dialog> JSX elements where a shadcn equivalent exists, and any local files that re-implement a primitive shadcn already ships.
The design system flows downstream from this commitment: token extraction emits shadcn-standard CSS variable names (--background, --foreground, --primary, --primary-foreground, --secondary, --muted, --accent, --destructive, --border, --input, --ring, --radius) in HSL channel form (222.2 47.4% 11.2%, not hsl(...)); brand-specific extensions (--brand-50 … --brand-900) are layered on top.
Docker for every infrastructure dependency — no local installs
Every infrastructure service the project needs runs in Docker via a single docker-compose.yml at the monorepo root. No local Postgres, no Homebrew Redis, no manually-installed Timescale extension. The repo clones, docker compose up -d starts the world, the apps connect.
Default services every project gets (set up by monorepo-bootstrapper in Phase B):
If any service is unhealthy, builders DO NOT run. The orchestrator surfaces the Docker status to the user, attempts docker compose up -d once, re-checks, and if still red, halts with a clear message.
Title Case for every enum value — DB = wire = UI label
Every enum/status/stage/role/discriminator on the wire uses Title Case. The string stored in Postgres equals the string the frontend renders. Zero conversion code.
Drizzle pgEnum values are Title Case strings (spaces allowed)
Zod z.enum([...]) literals in packages/contracts are Title Case
Discriminated-union kind fields are Title Case ({ kind: 'Email Sent', ... })
Numeric ranges ("1-10", "51-200") stay as ranges — not word-based, not converted
The label-map pattern (INDUSTRY_LABELS = { tech: 'Technology' }) is BANNED for simple enums — the value IS the label
Complex enums keep { kind, label } ONLY when label carries computed context (e.g., growth_signal.label = "+12% YoY"); kind is still Title Case
This convention is enforced in:
contracts-author agent when generating packages/contracts/src/*.ts
backend-module-builder when defining Drizzle pgEnum values
frontend-module-builder when authoring fixtures and rendering components
gap-auditor flags any PRD or DESIGN_DIGEST enum that uses non-Title-Case strings
plan-architect produces an entity catalog with Title Case enums baked in
integration-tester greps for casing helpers (capitalize, humanize, LABELS[, toLowerCase, toUpperCase) in component code
Prerequisites
This skill is the conductor. The two skills it conducts must be installed and available:
design-to-nextjs — frontend recipe
nestjs-enterprise-backend — backend recipe
Verify before starting:
ls ~/.claude/skills/design-to-nextjs/SKILL.md
ls ~/.claude/skills/nestjs-enterprise-backend/SKILL.md
If either is missing, install it before running this orchestrator.
The four-phase pipeline
Phase A: AUDIT Read-only. Extract from PRD + design, diff them, produce execution plan.
Phase B: BOOTSTRAP Sequential. Stand up the monorepo + shared contracts.
Phase C: EXECUTE Parallel per feature module. The speed lever.
Phase D: INTEGRATE Sequential. Run the cross-cutting validation tests.
Phase A ends with a mandatory user-confirmation gate — the orchestrator presents the EXECUTION_PLAN.md and waits for explicit approval before any code is written.
The agent team
Fifteen specialized subagents — ten core (build pipeline) and five security (loaded only if the security-review-and-fix skill is installed). Each has a focused role, a restricted tool set, and a single artifact it owns.
Full core-agent definitions live in references/agent-definitions.md. Full security-agent definitions live in the security-review-and-fix skill's references/security-agents.md. Full QA-agent definitions live in the exploratory-qa skill's references/qa-agents.md. The orchestrator's install step (Phase A.1) reads from each file when the relevant skill is detected.
Phase A — AUDIT
Goal: produce an EXECUTION_PLAN.md that lists every feature module to build, what each module needs (contracts, schemas, endpoints, screens), what the PRD/design disagree on, and what order to build in.
Step A.1 — Install agents
Before any subagent runs, the orchestrator copies all agent definitions into .claude/agents/. Always install the ten core agents from this skill. Also install the five security agents if and only if the security-review-and-fix skill is present (so Phase D.2 can dispatch them), and the four QA agents if and only if the exploratory-qa skill is present (so Phase D.3 can dispatch them).
mkdir -p .claude/agents
# 1) Install core agents from this skill
~/.claude/skills/prd-design-build-orchestrator/references/install-core-agents.sh
# 2) Conditionally install security agentsif [[ -f ~/.claude/skills/security-review-and-fix/references/security-agents.md ]]; then
~/.claude/skills/security-review-and-fix/references/install-security-agents.sh
echo"Security skill detected — security agents installed; Phase D.2 will run."elseecho"Security skill not detected — Phase D.2 will be skipped with a warning."fi# 3) Conditionally install exploratory QA agentsif [[ -f ~/.claude/skills/exploratory-qa/references/qa-agents.md ]]; then
~/.claude/skills/exploratory-qa/references/install-qa-agents.sh
echo"Exploratory QA skill detected — QA agents installed; Phase D.3 will run."elseecho"Exploratory QA skill not detected — Phase D.3 will be skipped with a warning."fils .claude/agents/
# Expected counts:# - 10 core agents always# - 15 with security skill installed# - 14 with QA skill installed# - 19 with both security and QA skills installed
Backend stack note:backend-module-builder (Nest.js) and laravel-module-builder (Laravel) are both auto-discovered plugin agents. The orchestrator dispatches whichever matches the backend_stack chosen at Step A.5b; the other simply goes unused. No conditional install step is needed — they ship with the plugin.
The full install scripts (with the awk-based extraction) live in:
references/agent-definitions.md of this skill (core agents)
references/security-agents.md of the security-review-and-fix skill (security agents)
references/qa-agents.md of the exploratory-qa skill (QA agents)
Run once at project start.
Step A.2 — Parallel inventory
The orchestrator (main session) dispatches prd-analyst and design-inventoryin parallel — both via the Agent tool in a single tool-use batch. In natural language:
"Use the prd-analyst subagent to read the PRD at <path> and produce PRD_DIGEST.md per the format in references/artifacts-format.md. In parallel, use the design-inventory subagent to inventory the design at <path> and produce DESIGN_DIGEST.md per the format in references/artifacts-format.md."
Claude Code emits both as Agent tool calls in one batch; they run concurrently with independent contexts. Wait for both to complete before advancing.
Note on tool naming: In Claude Code v2.1.63+, the tool for spawning subagents is named Agent. The previous name Task still works as an alias. The legacy Task(subagent_type='X', prompt='Y') shorthand used in older docs is equivalent to invoking Agent with the same parameters.
Step A.3 — Audit
The orchestrator dispatches gap-auditor:
"Use the gap-auditor subagent to read PRD_DIGEST.md and DESIGN_DIGEST.md and produce AUDIT.md."
The audit categorizes gaps as: missing-from-design, missing-from-prd, type-mismatch, naming-drift, scope-creep. See references/audit-pipeline.md.
Optional: agent-teams mode for the audit. For PRDs where the design diverges substantially or the cost of missed gaps is high, the gap audit benefits from adversarial review. If CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 is set, instead of dispatching a single gap-auditor subagent, ask the orchestrator:
"Spawn an agent team with three teammates using the gap-auditor agent type as a base: one PRD advocate (arguing the PRD is the source of truth), one design advocate (arguing the design is), and one neutral arbiter. Have them debate the gaps and produce AUDIT.md by consensus. Require the arbiter to sign off on each gap classification."
Token cost is roughly 3× the subagent path; recommended only for high-stakes audits.
Step A.4 — Plan
The orchestrator dispatches plan-architect:
"Use the plan-architect subagent to read PRD_DIGEST.md, DESIGN_DIGEST.md, and AUDIT.md, then produce EXECUTION_PLAN.md with build order, dependencies, and parallelizable lanes."
Step A.5 — User-confirmation gate
The orchestrator reads EXECUTION_PLAN.md, summarizes:
N features to build, grouped by parallel lane
Resolved gaps (PRD-only and design-only items now have decisions)
Estimated wave count (waves are batches of parallel feature builds)
Show this to the user. Do not proceed without explicit confirmation. If the user wants to revise (drop a feature, change an entity name, etc.), edit AUDIT.md and rerun plan-architect.
Step A.5b — Backend-stack selection gate
If EXECUTION_PLAN.md contains backend modules, the orchestrator asks the operator — before Phase B — which backend stack to build, using AskUserQuestion:
Persist the answer to STACK.md and a backend_stack: field in EXECUTION_PLAN.md so every later phase — and any resume — reads the same value. All backend routing in Phases B/C/D below is conditioned on backend_stack. When the backend is Nest.js, the frontend is always Next.js and the frontend half (design-to-nextjs, frontend-modular-architecture, QA/security/audit) is unaffected. When the backend is Laravel, the frontend stack is chosen separately at Step A.5c (Inertia monolith vs decoupled Next.js).
The agents this gate re-routes:
monorepo-bootstrapper (B.1) — Nest scaffold vs Laravel scaffold (see its stack-aware section).
contracts-author (B.2) — hand-authored Zod (Nest) vs Eloquent API Resources + hand-written TS (Laravel).
backend module builder (C.2) — backend-module-builder (Nest) vs laravel-module-builder (Laravel).
Phase D ship — add laravel-bref-deploy when the stack is Laravel.
If backend_stack == Laravel AND the plan has frontend modules, the orchestrator asks — before Phase B — which frontend to build, using AskUserQuestion:
Frontend for the Laravel backend?
Inertia monolith (default) — one Laravel app, React via Inertia (design-to-laravel); Fortify session auth; one Bref deploy.
Decoupled Next.js — separate apps/web (design-to-nextjs); Sanctum token auth; Laravel API + Next.js.
Persist frontend_stack to STACK.md / EXECUTION_PLAN.md. Routing when frontend_stack == Inertia:
bootstrap: a single Laravel app via the React starter kit (frontend in resources/js/; no apps/web, no pnpm web package) — see monorepo-bootstrapper.
auth:Fortify session + spatie/laravel-permission (see laravel-enterprise-backend/references/inertia-variant.md), NOT Sanctum tokens.
When frontend_stack == Next.js (with Laravel), use the decoupled path exactly as the backend-stack gate describes. When backend_stack == Nest.js, this gate does not run (frontend is always Next.js).
Phase B — BOOTSTRAP
Goal: monorepo skeleton + shared contracts in place, ready for parallel feature builds.
Step B.0 — Arm the Definition-of-Done gate (do this FIRST)
Before any code is written, arm the enforced done-gate so a premature "done" claim is impossible later:
From now on the Stop hook (hooks/scripts/done-gate.sh) blocks any turn that ends on a completion claim while any gate is red. See references/done-gate.md. Update COMPLETION_LEDGER.json as each wave gate and Phase D agent finishes (set the matching gate to pass, refresh head_sha). Do not delete the sentinel to silence the gate — fix the failing checks, or record a conscious override with .claude/.superdev-done-override.
Step B.1 — Monorepo bootstrap
Dispatch monorepo-bootstrapper:
"Use the monorepo-bootstrapper subagent to read EXECUTION_PLAN.md, then scaffold the pnpm workspace per nestjs-enterprise-backend/references/monorepo-setup.md, scaffold apps/api per nestjs-enterprise-backend/references/scaffolding.md, and scaffold apps/web per design-to-nextjs/references/scaffolding.md. Stop after pnpm install + first health check pass."
If backend_stack == Laravel (Step A.5b): instead scaffold apps/api as a Laravel 13 app per laravel-enterprise-backend/references/scaffolding.md + monorepo-setup.md (composer, Laravel Boost, stock pgsql → managed PostgreSQL + TimescaleDB, database cache/session tables, single-node Postgres+Timescale compose for local), and packages/contracts is hand-written TS kept in lockstep with the Eloquent API Resources (no codegen). See the stack-aware section in the monorepo-bootstrapper agent definition.
This is sequential and foundational — it must finish before contracts-author runs.
Step B.2 — Author all contracts up front
The orchestrator dispatches contracts-author:
"Use the contracts-author subagent to read EXECUTION_PLAN.md and, for each feature module, author the Zod schemas in packages/contracts/src/<feature>.ts following the view-shape contract in nestjs-enterprise-backend/references/view-presenter.md and the contracts patterns in nestjs-enterprise-backend/references/monorepo-setup.md."
If backend_stack == Laravel (Step A.5b): contracts-author instead authors Eloquent API Resources under apps/api/app/Domains/<feature>/Http/Resources/ plus the matching hand-written TypeScript in packages/contracts/src/<feature>.ts (decoupled Next.js) — or resources/js/types/ for the Inertia monolith — per laravel-enterprise-backend/references/api-resources.md. A Pest contract test locks each Resource to its TS shape. No spatie/laravel-data, no codegen. See the stack-aware section in the contracts-author agent definition.
Why all contracts up front? Because module builders in Phase C depend on the shared contracts (@<scope>/contracts for Nest and decoupled-Laravel; resources/js/types for the Inertia monolith) being complete. If contracts are written piecemeal alongside modules, the backend builder for module X can race the contracts for module Y.
After contracts-author finishes, run pnpm --filter @<scope>/contracts build (Nest) or confirm the hand-written packages/contracts TS type-checks and the Pest contract test passes (Laravel) before parallel builders depend on it.
Phase C — EXECUTE (the parallel phase)
Goal: all feature modules built, in both backend and frontend, as fast as possible.
Important architectural note: Subagents cannot spawn other subagents. The orchestrator (the main session running this skill) is the ONLY entity that dispatches subagents. Every wave dispatch, every wave-gate check, every retry is initiated by the main session. Builder subagents return results and stop; they do not chain.
Step C.1 — Read the lanes
EXECUTION_PLAN.md groups features into parallel waves. A wave is a batch of features that can be built concurrently because they have no inter-feature dependencies.
Example wave structure for an <APP_NAME>-like product:
Wave 1: auth, workspaces (foundation — everyone depends on these)
Wave 2: companies, contacts, mailboxes (parallel — no deps between)
Wave 3: campaigns, pipeline (depend on Wave 2)
Wave 4: ai, email, webhooks (depend on campaigns/pipeline)
Wave 5: analytics, audit (read-side, depend on everything)
Step C.2 — For each wave, the orchestrator dispatches builders in parallel
For each feature in the current wave, the orchestrator emits BOTH a backend builder AND a frontend builder in the same tool-use batch. The backend builder is backend-module-builder when backend_stack == Nest.js, or laravel-module-builder when backend_stack == Laravel (Step A.5b); the frontend builder is unchanged. In natural language to the main session:
"For Wave 2, dispatch six subagents in parallel: a backend-module-builder for companies, a frontend-module-builder for companies, a backend-module-builder for contacts, a frontend-module-builder for contacts, a backend-module-builder for mailboxes, a frontend-module-builder for mailboxes. Each gets the prompt 'Build the module per EXECUTION_PLAN.md feature: .' Wait for all to complete before advancing."
Claude Code translates this into six Agent tool calls in a single batch. They run concurrently with independent contexts. Wait for all six to complete before moving to the next wave.
Optional: agent-teams mode for pair-programming within a feature. For features where backend↔frontend contract negotiation is likely (e.g., a complex pipeline module where stages, scoring, and view shapes are still in flux), the user can ask the orchestrator to spawn each feature as a 2-teammate agent team instead of two independent subagents. Teammates can message each other ("frontend: I need lead.score.label in the view shape" → "backend: adding to presenter"). Requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. Token cost is roughly 1.5× the subagent path because teammates share more context. Recommended only for features flagged in EXECUTION_PLAN.md as having unresolved contract questions.
Step C.3 — Wave gate (orchestrator-driven)
After each wave, the orchestrator (not the builders) runs:
pnpm --filter @<scope>/api typecheck and pnpm --filter @<scope>/web typecheck — catch contract drift early
The full wave gate procedure in references/execution-pipeline.md (Docker health, lint, fixture validation, unit tests, UI audit)
If failures, the orchestrator dispatches a focused fixer subagent on the failing module (backend-module-builder or frontend-module-builder with a "fix the typecheck errors in module X" prompt)
Only advance to the next wave when all checks are green
The orchestrator can also wire PostToolUse hooks in .claude/settings.json to run typecheck automatically after every builder's Write or Edit finishes — see "Hook-driven wave gates" in references/execution-pipeline.md.
Step C.4 — Parallelism budget
Claude Code allows multiple Agent tool calls per turn, but each subagent consumes context and compute. Rules:
Cap concurrent subagents at 6 per tool-use batch. For waves with >3 features, the orchestrator runs two backend builders + one frontend builder, then the rest in a follow-up call.
Same wave, same call — never split a wave across multiple tool calls unless you hit the cap. The wave boundary is the synchronization point.
Independent waves can pipeline, but only if Wave N has no dependencies on Wave N-1's output. Default: don't pipeline; wait.
See references/execution-pipeline.md for batching details.
Phase D — INTEGRATE
Goal: prove the assembled system works end-to-end, and run a security audit before release.
Step D.1 — Run integration tests
The orchestrator dispatches integration-tester:
"Use the integration-tester subagent to run the full validation checklist: cross-workspace isolation test, CASL ability test (per role), audit decorator test (every mutation produces audit_logs row), view-shape contract test (no undefined in any view response), demo mode fixture validation, production mode boot. Report results."
Step D.2 — Security review (if security-review-and-fix skill is installed)
If the security-review-and-fix skill is available (verified at Step A.1), the orchestrator runs its six-phase audit. The five security agents are already in .claude/agents/, so the orchestrator dispatches them directly — do not start a separate orchestration session.
D.2.1 — Inventory (sequential, foundational):
The orchestrator dispatches:
"Use the security-inventory subagent to inventory the codebase per ~/.claude/skills/security-review-and-fix/references/inventory-checklist.md and produce SECURITY_INVENTORY.md."
These three share no inputs; they all append to SECURITY_FINDINGS.md using disjoint ID prefixes (S-S-*, S-D-*, S-P-*), so race-safe. The orchestrator dispatches all three in a single batch:
"Dispatch three subagents in parallel:
Use the static-auditor subagent to run the full static audit per static-audit-checklist.md, read SECURITY_INVENTORY.md for context, and append findings to SECURITY_FINDINGS.md with prefix S-S-.
Use the dynamic-auditor subagent to verify the stack is up (docker compose ps + curl /readiness), then run probes per dynamic-audit-checklist.md and append findings with prefix S-D-.
Use the dependency-auditor subagent to run pnpm audit and lockfile checks per dependency-audit-checklist.md and append findings with prefix S-P-."
Wait for all three.
Optional: agent-teams mode for the audit phase. Security audits are a strong fit for the "competing hypotheses" pattern. Instead of three independent subagents, three teammates that can challenge each other's findings often surface more issues. If CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1:
"Spawn a 3-teammate agent team using the static-auditor, dynamic-auditor, and dependency-auditor agent types as bases. Have them debate findings as they go (a static-flagged SQL injection should be cross-checked against the dependency auditor's CVE database; a dynamic auditor's CSRF observation should be validated against static analysis of the upstream middleware). Each teammate's findings are challenged before being recorded. Produce SECURITY_FINDINGS.md by consensus."
The user can directly message any auditor mid-investigation via Shift+Down. Recommended for pre-launch audits and post-incident reviews. The security-inventory and security-fixer agents remain subagents in both modes.
D.2.5 — Triage (orchestrator + user, no subagent):
The orchestrator reads SECURITY_FINDINGS.md, walks the user through Critical → High → Medium following the triage prompts in ~/.claude/skills/security-review-and-fix/references/fix-plan-format.md, and writes SECURITY_FIX_PLAN.md. This is not a subagent dispatch — it's interactive with the user.
D.2.6 — Apply fixes (per-fix dispatches):
For each fix item in SECURITY_FIX_PLAN.md that has Agent: security-fixer, the orchestrator dispatches one subagent per fix (never batch):
"Use the security-fixer subagent to apply fix F-2 per SECURITY_FIX_PLAN.md. Edit only the files listed in that entry. Run pnpm typecheck and the relevant tests. Re-grep for the original anti-pattern to confirm it is gone. Return per the format in security-agents.md."
Run sequentially through the Critical/High fixes. Same-module Mediums can batch into a single dispatch only when the plan explicitly lists them as a group.
D.2.7 — Re-audit:
After all fixes applied, re-run the three auditors in parallel (same prompts as D.2.2–D.2.4). Append a "RE-AUDIT" section to SECURITY_FINDINGS.md. The originally-flagged patterns should no longer match; any that do represent incomplete fixes and trigger another fix pass (max 2 re-audit cycles before escalating to user).
Halt conditions for D.2:
Critical findings unresolved after Phase D.2.7 → halt; do not produce the final report. Surface SECURITY_FIX_PLAN.md with the unresolved items to the user.
High findings with explicit user deferrals → record in the final report's "accepted risks" section.
Medium/Low/Info → record as recommendations in the final report.
If the security skill is not installed (Step A.1 didn't install the security agents), skip D.2 entirely and add to the final report: "⚠ Pre-launch security review was not performed. Install the security-review-and-fix skill and re-run before production deployment."
Step D.3 — Exploratory QA (if exploratory-qa skill is installed)
If the exploratory-qa skill is available, run it after security passes. Exploratory QA finds the issues that automated tests systematically miss (missing empty states, button-size drift across modules, frozen frames on large data, refactor candidates) — different category from functional regressions and security.
Install the QA agents if not already done (Phase A.1 should have detected and installed them):
if [[ ! -f .claude/agents/qa-flow-tester.md ]]; then
~/.claude/skills/exploratory-qa/references/install-qa-agents.sh
fi
Then run the six-phase QA pipeline. Same dispatch pattern as the security skill — agents are local to the project's .claude/agents/, the orchestrator drives them.
D.3.1 — Environment (sequential):
The orchestrator dispatches:
"Use the qa-environment subagent to verify stack health, seed at realistic scale per ~/.claude/skills/exploratory-qa/references/environment-checklist.md, capture baseline screenshots, plant edge-case fixtures, and capture test credentials. Produce QA_ENVIRONMENT.md."
D.3.2 — Happy path (parallel per feature, batched at 6):
The orchestrator dispatches one qa-flow-tester per feature in a single tool-use batch (capped at 6):
"Dispatch six qa-flow-tester subagents in parallel, each scoped to one feature module (companies, contacts, mailboxes, campaigns, pipeline, mailboxes), mode happy-path, following the recipes in flow-recipes.md. Each writes to qa/flows/<feature>/happy-path/."
For projects with more than 6 features, run in successive batches; do not exceed 6 concurrent subagents per tool-use turn.
D.3.3 — Edge cases (parallel per feature, same batching):
The orchestrator dispatches:
"Dispatch qa-flow-tester subagents in parallel (max 6 per batch), each scoped to one feature, mode edge-cases. Cover all 13 categories: empty-state, loading-state, error-state, large-data, slow-network, validation-errors, concurrent-mutation, stale-data, long-content, special-characters, keyboard-nav, mobile, accessibility."
D.3.4 — Consistency (single subagent, runs after flows finish):
"Use the qa-consistency-checker subagent to audit cross-module consistency per consistency-checklist.md. Read all qa/flows/*/observations.md and the frontend source. Produce QA_CONSISTENCY.md."
D.3.5 — Performance (single subagent, runs LAST with idle system):
"Use the qa-performance-prober subagent to measure Web Vitals per route, database query counts per action, frontend data-volume freezes, bundle sizes, worker latency, and memory growth. Run only after all flow-testers complete and the system is idle. Produce QA_PERFORMANCE.md."
Optional: agent-teams mode for adversarial report synthesis. Phase D.3.6 (final QA report synthesis) can run as a 3-teammate agent team that debates severity assignment before writing the report. Useful when the report drives launch/no-launch decisions. If CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1:
"Spawn a 3-teammate agent team for the QA synthesis: a harshest-critic teammate pushing for higher severities, a pragmatist teammate pushing back, and a shipping-advocate teammate arguing what genuinely blocks launch. Have them debate each finding's severity. Produce QA_REPORT.md by consensus."
Optional: competing-hypotheses team for hard performance issues. When qa-performance-prober flags an issue with multiple plausible causes (e.g., "companies list LCP 3.8s — could be N+1, bundle size, or unindexed query"), the orchestrator can on-demand spawn a 3-teammate investigation team (one investigator per hypothesis) that messages each other to disprove competing theories. Same pattern as the docs' "scientific debate" example.
D.3.6 — Synthesis (orchestrator, no subagent):
The orchestrator reads QA_ENVIRONMENT.md, all qa/flows/*/observations.md, QA_CONSISTENCY.md, QA_PERFORMANCE.md and writes QA_REPORT.md per ~/.claude/skills/exploratory-qa/references/qa-report-format.md.
Halt conditions for D.3:
Critical findings unresolved → halt; surface QA_REPORT.md to the user before producing the orchestrator's final report.
High findings need explicit user acknowledgment in the final report's "accepted risks" section.
Medium/Low/Refactor are recorded as recommendations.
If the QA skill is not installed, skip D.3 with: "⚠ Exploratory QA was not performed. Install the exploratory-qa skill and re-run for a production-readiness pass."
Step D.3.5 — Production readiness loop (the capstone — rate every module to 9+)
After integration, security, and QA have run, invoke production-readiness-audit as the capstone. This is the scored, iterative loop that turns "looks done" into "is done" — and it's what satisfies the user's recurring ask to "rate each module 1–10 and complete it fully for production." It subsumes product-completeness-audit and brutal-exhaustive-audit into one scored loop.
Precondition: verify the stack is up in production mode (demo OFF) and dispatch seed-data-architect so there is real seed data + an initial owner/admin login + a second tenant (for the tenancy test).
Five passes score every module 0–10: (1) completeness & wiring, (2) data realism & seed, (3) tenancy isolation, (4) end-to-end via Playwright — every route, every form, every data shape FE↔BE, (5) code quality & security (zero lint/type errors, no suppressions). module_score = min(passes); a tenancy leak or P0 caps a module at ≤3.
Rate:module-readiness-rater writes READINESS.md + READINESS_CHECKLIST.md and updates COMPLETION_LEDGER.json (features.<module>.readiness_score, gates.readiness).
Fix & loop: for every module < 9, dispatch the matching fixer (backend/frontend/laravel/inertia-module-builder, security-fixer) on its checklist items — root cause only, no suppressions, no new demo data — then re-run the sub-9 passes. Repeat until every module ≥ 9 or round 5 (then surface the remaining blockers to the user; do not lower the bar).
See production-readiness-audit. The done-gate will not pass until gates.readiness == pass and every readiness_score ≥ 9, so this loop is the only legitimate path to a "done" claim.
Step D.4 — Definition-of-Done gate, THEN final report
Phase C green is not done. Before writing the final report or claiming completion, every Phase D agent must have run and recorded its verdict in COMPLETION_LEDGER.json, and the gate must pass:
# update the ledger gates as each Phase D agent finished, then verify:
bash "$CLAUDE_PLUGIN_DIR/hooks/scripts/done-gate.sh" --report # exit 0 = DONE# (or invoke the /superdev-done skill)
The gate (see references/done-gate.md) requires: all ledger gates pass (typecheck, lint, build, integration, completeness, security, qa, brutal), a fresh ledger (head_sha == HEAD), zero new suppressions since base_sha, and zero demo/placeholder content in apps/web/src. The Stop hook enforces this automatically — a "done" claim while any check is red is blocked.
Only once the gate is green, the orchestrator collects and presents:
Wave-by-wave build status (from Phase C)
Test report (from integration-tester)
Security review summary (from D.2, if performed)
Completeness + QA + brutal-audit verdicts (from the ledger)
Any explicitly deferred items (accepted risks recorded in the ledger)
Files created / lines of code summary
Present to the user. Done — and provably so.
Reference files
File
When to read
references/orchestration-patterns.md
Before any subagent dispatch — the Task tool + parallelism rules
references/audit-pipeline.md
Phase A — what each audit agent does and how findings are categorized
references/execution-pipeline.md
Phase C — wave construction, batching, error handling
references/artifacts-format.md
Reading or writing PRD_DIGEST / DESIGN_DIGEST / AUDIT / EXECUTION_PLAN
references/agent-definitions.md
Phase A.1 — install agents to .claude/agents/. Source-of-truth for every subagent.
references/done-gate.md
Phase B.0 + Phase D.4 — the enforced Definition-of-Done gate, the COMPLETION_LEDGER.json schema, and what arms/overrides it.
Validation checklist
Before declaring the build done:
done-gate.sh --report (or /superdev-done) exits 0 — this is the single overriding gate; every item below is also encoded in it. A "done" claim while it is red is blocked by the Stop hook.
COMPLETION_LEDGER.json exists with every gate pass (typecheck, lint, build, integration, completeness, security, qa, brutal, readiness) and head_sha == current HEAD
Every module scores ≥ 9 in production-readiness-audit (features.*.readiness_score ≥ 9); seed data + initial owner/admin user present; tenancy isolation proven; every route + every form exercised in Playwright (FE↔BE shapes match)
Every Phase D agent actually ran — integration-tester, product-completeness-audit, security-review-and-fix, exploratory-qa, brutal-exhaustive-audit, production-readiness-audit (not skipped, not deferred-without-acknowledgment)
Zero new suppressions since base_sha — no eslint-disable / @ts-ignore / @ts-expect-error / as any / as unknown as
Zero demo/placeholder content in apps/web/src (no coming soon, lorem ipsum, hardcoded test cards, mockData/fakeUsers outside mocks/tests)
.claude/agents/ contains all 10 core agent definitions
If security skill is installed, .claude/agents/ also contains the 5 security agents (15 total)
If exploratory-qa skill is installed, .claude/agents/ also contains the 4 QA agents (14 with QA only; 19 with security and QA)
PRD_DIGEST.md, DESIGN_DIGEST.md, AUDIT.md, EXECUTION_PLAN.md exist at project root
If security skill ran in D.2: SECURITY_INVENTORY.md, SECURITY_FINDINGS.md, SECURITY_FIX_PLAN.md exist; zero unresolved Critical findings
If QA skill ran in D.3: QA_ENVIRONMENT.md, QA_CONSISTENCY.md, QA_PERFORMANCE.md, QA_REPORT.md exist; zero unresolved Critical QA findings
User confirmed EXECUTION_PLAN.md before Phase B started
docker-compose.yml exists at the monorepo root (not inside apps/api/)
docker compose up -d brings every service to healthy state (postgres, redis, and any plan-specific extras)
No infrastructure dependency is installed locally — Postgres, Redis, Timescale, SMTP, S3 all run in Docker
Root package.json has dev:infra, dev:infra:down, dev:infra:reset, dev:infra:logs scripts
Root .env.example documents every Docker-service connection string
shadcn/ui initialized in apps/web with all primitives installed (components.json present, src/components/ui/ populated, sidebar block installed)
ui-auditor ran clean at the final wave gate — no forbidden UI library imports, no raw <button>/<input>/<dialog> JSX where shadcn equivalent exists, every sidebar uses shadcn's sidebar block
pnpm install succeeds at the monorepo root
pnpm --filter @<scope>/contracts build succeeds
pnpm turbo build succeeds across all packages
pnpm turbo typecheck is green
pnpm turbo lint is zero-warning
All integration tests from Phase D pass
Title Case enum sweep clean — grep -rEn "(capitalize|humanize|\\.toLowerCase\\(\\)|\\.toUpperCase\\(\\)|_LABELS\\[|LABELS\\[)" apps/web/src/modules returns zero hits on contract-typed data; every Drizzle pgEnum uses Title Case values
Final report has zero unresolved blocking audit findings
No agent definition was loaded that wasn't in the install list (no rogue agents)
Common pitfalls
P1 — Skipping the user-confirmation gate. The orchestrator must show EXECUTION_PLAN.md and wait for explicit "yes, proceed". A subtly wrong plan costs hours of wasted parallel builds.
P2 — Authoring contracts inside the module builders. Then two features race on packages/contracts. Always do all contracts in Phase B.2.
P3 — Pipelining waves. Tempting for speed; lethal for correctness when Wave N depends on Wave N-1's types. Default: wait at wave boundaries.
P4 — Letting subagents talk to each other. They can't. State lives on disk. If two agents need to coordinate, that's a sign the orchestrator should split the work differently.
P5 — Giving every agent every tool. A read-only agent (e.g. prd-analyst) given Write/Bash will sometimes "helpfully" create files prematurely. Restrict tools per agent.
P6 — One mega-prompt to one mega-agent. That's not multi-agent; that's a long-context single agent. The win comes from specialization + parallel execution, not from delegation alone.
P7 — Trusting agent self-reports. A backend-module-builder may report "done" with a broken typecheck. Always run the wave-gate typecheck (Step C.3) before advancing.
P8 — Forgetting the design-to-nextjs / nestjs-enterprise-backend skill references. The builders need them. Verify the skills are installed at Phase A start.
