web-architecture

Multi-agent orchestration for complex TypeScript/Next.js/Convex projects. Phased builds, functional verification, the full playbook for delegating to sub-agents without chaos.

Safety Notice

This listing is from the official public ClawHub registry. Review SKILL.md and referenced scripts before running.

Copy this and send it to your AI assistant to learn

Install skill "web-architecture" with this command: npx skills add michaelmonetized/web-architecture

Web Architecture

Multi-agent development workflow for TypeScript/Next.js/Convex projects.

Born from: 29 agents, 50K lines, 212 errors, 1 hard lesson


⚠️ SUB-AGENT POLICY (READ FIRST)

No Timeouts

Sub-agents MUST run without timeout limits. A 10-minute timeout that cuts off an agent mid-implementation leaves broken, partial code. Let agents finish.

Completion Means Working, Not Compiling

"Build passes" is necessary but NOT sufficient.

Before marking ANY phase complete, verify:

  1. Functions actually work — Call them, verify data flows
  2. UI actually renders data — Not just loading spinners forever
  3. User flows complete end-to-end — Click through, verify state changes persist
  4. Error states are handled — Not just happy path

The Lesson

An agent produced 15K lines of "working" code that:

  • ✅ Compiled with zero TypeScript errors
  • ✅ Passed bun run build
  • ❌ Had ZERO actual functionality
  • ❌ All data was mocked or hardcoded
  • ❌ Every button was a no-op

Self-grade: 5/10 — A prototype, not a product.


The Core Lesson

Single agent with full context > Many agents with partial context

29 parallel agents wrote 50K lines of code that didn't compile. Why?

  • No schema coordination → duplicate table definitions
  • No type contracts → frontend expected user.role, backend returned profile.plan
  • No initialization → npx convex dev never ran, no generated types
  • No integration checkpoints → errors discovered only at the end

The fix: One agent with full context rewrote the entire Convex backend in 11 minutes.


When to Use Multi-Agent

Good for parallel work:

  • Marketing pages (after design system exists)
  • Documentation files (independent)
  • Isolated features with clear contracts

Bad for parallel work:

  • Schema design (needs single owner)
  • Core type definitions (must be shared)
  • Interconnected backend functions
  • Component library (needs consistency)

The Workflow

Phase 0: Bootstrap (SEQUENTIAL — One Agent)

Must complete before spawning ANY other agents.

  1. Initialize project structure
  2. Initialize Convex: npx convex dev --once
  3. Create complete schema.ts (ALL tables)
  4. Run npx convex dev to generate types
  5. Create CONTRACTS.md (all data shapes)
  6. Create shared types in lib/types.ts
  7. Verify: bun run build passes

Deliverables:

  • convex/schema.ts — Complete, no TODOs
  • convex/_generated/ — Types generated
  • CONTRACTS.md — API shapes documented
  • lib/types.ts — Shared frontend types
  • bun run build — Passes with 0 errors

Phase 1: Foundation Documents (CAN BE PARALLEL)

Only spawn AFTER Phase 0 completes.

AgentOutputDependencies
Tech RequirementsTECH-REQ.mdNone
ComplianceCOMPLIANCE.mdNone
Design PrinciplesDESIGN.mdNone
Coding StandardsSTANDARDS.mdNone

Rule: These agents READ the schema. They do NOT modify it.


Phase 2: Backend Implementation (SEQUENTIAL or CAREFUL PARALLEL)

Option A: Single Backend Agent (Recommended)

  • One agent implements all Convex functions
  • Consistent patterns, no conflicts

Option B: Parallel with File Locks

  • Each agent owns specific files
  • NO shared file writes
  • Must reference CONTRACTS.md

Functional Requirements:

  1. Test CRUD operations — Create, read, update, delete
  2. Verify queries return data — Not empty arrays
  3. Check mutations persist — Data survives refresh
  4. Test auth guards — Protected functions reject unauthorized
  5. Verify indexes work — Queries return correct filtered data

Phase 3: Component Library (SEQUENTIAL)

Single agent builds the component library.

Why? Components reference each other. Parallel work creates duplicate components with different APIs.

Functional Requirements:

  1. Interactive states work — Buttons trigger onClick
  2. Form components submit — Not just styled divs
  3. Loading/error states exist
  4. Accessibility basics — Labels, ARIA, keyboard nav
  5. Consistent API — All components follow same patterns

Phase 4: Features & Pages (CAN BE PARALLEL)

Now safe to parallelize because schema is locked, types exist, components exist.

AgentScopeCan Modify
Admin Suite/app/(admin)/**Own files only
Support Portal/app/(support)/**Own files only
Marketing Pages/app/(marketing)/**Own files only
User Flows/app/(app)/**Own files only

Rules:

  1. Read schema, types, contracts — don't modify
  2. Use existing components — don't recreate
  3. Write to assigned directories only

Functional Requirements:

  • Page loads without console errors
  • Data appears (not mock/placeholder)
  • Forms submit and persist data
  • Can complete full user flow (create → view → edit → delete)
  • Refresh preserves state

Red flags (NOT complete):

  • // TODO comments in business logic
  • Hardcoded arrays instead of useQuery
  • onClick handlers that console.log instead of mutate
  • "Coming soon" placeholders in core features

Phase 5: Integration & QA (SEQUENTIAL)

  1. bun run build (must pass)
  2. npx convex dev --once (must pass)
  3. Generate sitemap from routes
  4. Route crawl & 404 check
  5. Browser smoke test (all routes return 200)
  6. End-to-end flow verification

E2E Verification Checklist:

Auth Flow:

  • Sign up creates user in database
  • Sign in authenticates and redirects
  • Protected routes redirect to sign-in

Core CRUD Flow:

  • Create: Form submits → record appears
  • Read: List shows real data
  • Update: Edit form saves → changes persist
  • Delete: Remove action → record gone

Directory Structure

project/
├── convex/
│   ├── schema.ts            # 🔒 Phase 0 only
│   ├── _generated/          # 🔒 Auto-generated
│   └── [domain].ts
├── lib/
│   ├── types.ts             # 🔒 Phase 0 only
│   └── utils.ts
├── components/
│   ├── ui/                  # Component library agent
│   └── [domain]/            # Feature agents
├── app/
│   ├── (admin)/             # Admin agent
│   ├── (app)/               # App agent
│   └── (marketing)/         # Marketing agents
└── CONTRACTS.md             # 🔒 Phase 0 only

🔒 = Locked after Phase 0. Agents read, don't modify.


Agent Spawn Order

1. Bootstrap Agent (MUST COMPLETE FIRST)
   └── schema.ts, types, contracts
   
2. Doc Agents (parallel)
   ├── TECH-REQ.md
   ├── COMPLIANCE.md
   └── DESIGN.md
   
3. Backend Agent (single)
   └── All convex/*.ts functions
   
4. Component Agent (single)
   └── All components/ui/*
   
5. Feature Agents (parallel, isolated directories)
   ├── Admin Suite
   ├── Support Portal
   ├── Marketing Pages
   └── User Flows
   
6. Integration Agent (single)
   └── Final build, fixes, QA

Anti-Patterns

Spawn all agents at once — No coordination, duplicate work

Let agents invent types — Use CONTRACTS.md, not imagination

Skip Phase 0 — "We'll figure out the schema later" = disaster

Parallel schema writes — One owner only

Frontend before backend types — Generates type mismatches

No build checkpoints — Errors compound


Related Files

Source Transparency

This detail page is rendered from real SKILL.md content. Trust labels are metadata-based hints, not a safety guarantee.

Related Skills

Related by shared tags or category signals.

Coding

A Python CLI skill for Cutout.Pro visual APIs — background removal, face cutout, and photo enhancement. Supports file upload & image URL input.

Call Cutout.Pro visual processing APIs to perform background removal, face cutout, and photo enhancement. Supports both file upload and image URL input, retu...

Registry SourceRecently Updated
Coding

client-onboarding-agent

Client onboarding and business diagnostic framework for AI agent deployments. Covers 4-round diagnostic process, 6 constraint categories, deployment SOP with...

Registry SourceRecently Updated
Coding

Ai Tools

AI Tools Box - Search and invoke 100+ AI tools. Categories: Writing, Image, Video, Coding, Office, Search, Chat, Audio, Design, Agent, Translation, Dev Platf...

Registry SourceRecently Updated