engineering-rapid-prototyper（工程快速原型机）

Rapid Prototyping Guide

Overview

This guide covers fast proof-of-concept development and MVP creation. Use it when building prototypes to validate hypotheses, setting up A/B testing, or choosing rapid development stacks that prioritize speed-to-deploy over production hardening.

Stack Selection Guide

How to choose tools

• When authentication is needed, use Clerk or NextAuth for instant setup — never build custom auth for a prototype. If the prototype does not need user accounts to test the hypothesis, skip auth entirely.
• When a database is needed, use Prisma + Supabase for instant hosting, schema management, and row-level security. If the prototype only needs to store <100 records for testing, use a JSON file or localStorage — do not set up a database.
• When deployment is needed, use Vercel for instant hosting and preview URLs on every PR. First deploy must happen within 2 hours of project start.
• When no-code/low-code can cover the requirement, use it — speed to validation matters more than custom code.
• When building, implement core functionality first, polish and edge cases later. If you are writing error handling before the happy path works end-to-end, stop.

Decision Rules for Scope

• Maximum 5 database tables. If the schema needs >5 tables, you are overbuilding — cut scope or simplify the data model.
• Maximum 3-5 features. If the feature list exceeds 5, rank by hypothesis impact and cut the bottom ones. A prototype tests one hypothesis well, not five hypotheses poorly.
• Maximum 8 pages/routes. If more are needed, combine views or cut features.
• Zero custom UI components — use shadcn/ui, Radix, or similar. If a component does not exist in the library, simplify the design to fit what exists.
• No custom backend logic beyond CRUD + one business rule. If the prototype requires complex backend processing (queues, scheduled jobs, multi-step workflows), mock it with a simple endpoint that returns realistic fake data.
• No optimization. No caching, no CDN config, no database indexes, no lazy loading, no code splitting. These are production concerns. The prototype must be fast enough to demo — nothing more.

Kill Criteria

• If setup (project init → first deploy) takes >3 hours, the stack is too complex. Switch to a simpler approach.
• If after 2 days the core hypothesis cannot be tested in the prototype, redesign the prototype — not the hypothesis.
• If a single feature takes >1 day to implement, the feature is too complex for a prototype. Simplify or fake it with static data.
• If you are writing tests, you have left prototype mode. Prototypes are disposable. Tests are for code you intend to keep.

Recommended Stack

• Framework: Next.js 14+ (App Router)
• UI: shadcn/ui + react-hook-form + Zod validation
• State: Zustand (only if client state is needed — prefer server components + server actions)
• Database: Prisma + Supabase (only if persistence is needed to test the hypothesis)
• Auth: Clerk or NextAuth (only if user identity is part of the hypothesis)
• Animation: Framer Motion (only for demos — skip if internal testing only)
• Deploy: Vercel

See Stack Setup for the full package.json and shadcn/ui install commands.

Workflow

Hour 0-2: Hypothesis and Setup

• Write the hypothesis in one sentence: "We believe [user] will [action] because [reason], measured by [metric]." If you cannot write this sentence, you are not ready to build.
• Scaffold project with scripts/scaffold.sh or npx create-next-app. Install shadcn/ui. Deploy empty shell to Vercel. This must be done in <2 hours.
• If auth is needed, configure Clerk (15 min). If database is needed, create Supabase project and run prisma db push (20 min). If neither is needed for the hypothesis, skip both.

Hour 2-8: Core Flow (Day 1)

• Build the single most important user flow end-to-end: entry point → core action → visible result. No side features, no settings, no edge cases.
• Use server components + server actions for data fetching. Only add client components when interactivity requires it.
• Hardcode anything that is not part of the hypothesis. Use seed data, placeholder images, and realistic fake content. If a feature needs an external API that is not set up, return mock data from a server action.
• Deploy to Vercel at end of day 1. The prototype must be clickable by a real user on day 1.

Hour 8-20: Supporting Features (Day 2-3)

• Add 2-4 supporting features, one at a time. Deploy after each feature. If a feature breaks the deploy, revert it immediately — do not debug for >30 minutes.
• Add a feedback mechanism: either a simple form (shadcn/ui textarea + server action that writes to Supabase) or a Hotjar/PostHog snippet (3 lines of code).
• If A/B testing is needed, use Vercel Edge Config or a simple cookie-based split — not a full experimentation platform.

Hour 20-24: Test and Decide (Day 3-4)

• Share with 5-10 target users. Observe or collect feedback for 24-48 hours.
• Measure against the hypothesis metric defined in hour 0. If the metric is positive, plan production build. If negative, decide: pivot the hypothesis or kill the prototype.
• Never iterate on a prototype for >1 week. After 1 week, either commit to production-quality rebuild or archive it.

See Code Examples for a feedback form component and A/B testing hook.

Validation Feedback Loop

The prototype exists to answer a question. Use this protocol to ensure it actually does:

• Before building: Write the hypothesis, the metric, and the threshold for success/failure. Example: "If >30% of test users complete the checkout flow without help, the UX hypothesis is validated."
• During building: After each feature, deploy and check if the feature is testable by a real user. If it is not clickable end-to-end, it is not done.
• Day 1 checkpoint: Share the deployed prototype with 1 person (yourself or a teammate). Can they complete the core flow without explanation? If not, the flow is broken — fix before adding features.
• Day 2-3 checkpoint: Share with 5-10 target users. Observe or collect feedback. If users are confused by the same step, that is the blocker — fix that one step, do not add more features.
• Decision point (day 3-5): Measure the hypothesis metric. Three outcomes:

- Metric met: Hypothesis validated. Document the learnings, archive the prototype, begin production planning with the production skills. - Metric close but not met: One iteration allowed. Identify the single biggest friction point, fix it, re-test with 5 more users. If still not met, kill it. - Metric far from target: Hypothesis rejected. Document why. Decide: pivot (new hypothesis, new prototype) or kill (move to a different idea entirely).

Self-Verification Protocol

Prototypes do not need test suites, but they need basic verification:

• The app deploys successfully to Vercel (or equivalent) from the main branch. If deploy is broken, nothing else matters.
• The core flow works end-to-end: a new user can complete the primary action without hitting an error screen or dead end.
• The feedback mechanism works: submit test feedback and verify it is captured (check Supabase table, Hotjar recording, or PostHog event).
• The A/B split (if any) is working: verify both variants render by toggling the split cookie/flag.
• No placeholder text ("Lorem ipsum", "TODO", "test") is visible to test users — it undermines credibility and skews feedback.
• The prototype loads in <3 seconds on a mobile connection (throttle in DevTools to "Fast 3G" and verify). Slow prototypes get abandoned before the hypothesis is tested.

Scripts

• scripts/scaffold.sh -- Create a Next.js + TypeScript + Tailwind project structure with minimal boilerplate files (no npm install). Run with --help for options.

Anti-Patterns (Stop Immediately If You Catch Yourself Doing These)

• Writing unit tests or integration tests for prototype code.
• Setting up CI/CD pipelines, linting rules, or pre-commit hooks.
• Creating a design system, custom theme, or component library.
• Building an admin panel, settings page, or user profile page (unless it is the hypothesis).
• Adding error boundaries, retry logic, loading skeletons, or empty states.
• Refactoring code for "cleanliness" — prototype code is disposable.
• Debating technology choices for >15 minutes — pick the default stack and move on.
• Building features "while we are at it" that are not in the 3-5 feature list.

Transition to Production

When the hypothesis is validated and the decision is to build for real:

• Do NOT refactor the prototype into production code. Start fresh with the production skill (engineering-frontend-developer, engineering-backend-architect, etc.).
• Extract from the prototype: validated user flows, data model shape, API contract, and UI screenshots. These are the spec for the production build.
• Archive the prototype repo. It served its purpose.

References

• Stack Setup -- Package.json and shadcn/ui install commands.
• Code Examples -- Feedback form component and A/B testing hook.
• Full Stack Integration -- Prisma schema, Supabase + Clerk auth, server actions, tRPC, file uploads, and email with Resend.
• UI Patterns -- Data tables, command palette, form wizard, dashboard layout, toasts, and loading skeletons with shadcn/ui.