afrexai-claude-code-production（afrexai Claude 代码 production）

Claude Code Production Engineering

The complete methodology for shipping production code with Claude Code at 10X speed. Not installation scripts — actual patterns, workflows, and techniques that compound your output.

Quick Health Check (run mentally before every session)

Score: /8. Below 6 = slow, buggy sessions. Fix before coding.

1. Project Setup — CLAUDE.md Architecture

CLAUDE.md is your project's brain. Claude reads it at session start. A good one saves thousands of tokens per session.

Template

# Project: [name]

## Tech Stack
- Language: TypeScript (strict mode)
- Framework: Next.js 15 App Router
- Database: PostgreSQL via Drizzle ORM
- Testing: Vitest + Playwright
- Styling: Tailwind CSS v4

## Architecture Rules
- Max 50 lines per function, 300 lines per file
- One responsibility per file
- All exports typed — no `any`
- Errors as values (Result type), not thrown exceptions
- Database: migrations via `drizzle-kit generate` then `drizzle-kit push`

## File Structure
src/
  app/         → Next.js routes (thin — call services)
  lib/         → Business logic (pure functions)
  db/          → Schema, migrations, queries
  components/  → UI (server components default, 'use client' only when needed)
  types/       → Shared type definitions

## Commands
- `pnpm dev` — start dev server
- `pnpm test` — run vitest
- `pnpm test:e2e` — run playwright
- `pnpm lint` — eslint + tsc --noEmit
- `pnpm db:generate` — generate migration
- `pnpm db:push` — apply migration

## Conventions
- Imports: absolute from `@/` (mapped to `src/`)
- Naming: camelCase functions, PascalCase components/types, SCREAMING_SNAKE constants
- Commits: conventional commits (feat:, fix:, refactor:, test:, docs:)
- PRs: always create branch, never commit to main directly

CLAUDE.md Rules

1. Be specific — "TypeScript strict" not "use types." Stack versions, not just names.
2. Include commands — Claude needs to know how to run things. Exact commands, not descriptions.
3. Architecture decisions — document WHY, not just what. "Errors as values because we use Result type" tells Claude the pattern.
4. Keep it under 200 lines — CLAUDE.md is read every session. Bloat wastes tokens.
5. Update when patterns change — stale CLAUDE.md causes Claude to fight your codebase.
6. Nested CLAUDE.md — subdirectories can have their own. Claude merges them. Use for monorepo packages.

.claueignore

node_modules/
.next/
dist/
coverage/
*.lock
.git/
*.min.js
*.map
public/assets/

Rule: if Claude doesn't need to read it, ignore it. Large lock files and build artifacts waste context.

2. Prompting Patterns — The 5 That Matter

Pattern 1: Task Brief (use for any non-trivial work)

Task: Add user authentication with magic links
Context: Using Resend for email, no password system exists yet
Constraints:
- Server actions only (no API routes)
- Session via httpOnly cookies
- Token expires in 15 minutes
Acceptance: User enters email → receives link → clicks → logged in → cookie set
Start with: the database schema for sessions and tokens

Why it works: scope + constraints + acceptance criteria + starting point. Claude doesn't wander.

Pattern 2: Show, Don't Tell

Bad: "Make the API more robust" Good: "Add input validation to POST /api/users — validate email format, name 1-100 chars, reject extra fields. Return 422 with field-level errors matching this shape: { errors: { field: string, message: string }[] }"

Rule: if you can't describe the exact output shape, you don't know what you want yet. Think first.

Pattern 3: Incremental Refinement

Step 1: "Create the database schema for a todo app with projects and tasks"
[review output]
Step 2: "Now add the CRUD service layer for tasks — pure functions, no framework imports"
[review output]
Step 3: "Now the API routes that call those services — input validation with zod"

Why: Claude produces better code in focused steps than in one massive prompt. Each step builds verified context.

Pattern 4: Fix With Evidence

Bad: "It's broken" Good: "Running pnpm test gives this error:

TypeError: Cannot read properties of undefined (reading 'id')
  at getUserById (src/lib/users.ts:23:15)

The function expects a User object but receives undefined when the database query returns no rows. Add a null check and return a Result type."

Rule: paste the actual error. Claude is excellent at fixing bugs when it can see the stack trace.

Pattern 5: Architecture Discussion

I'm deciding between these approaches for real-time updates:
A) Server-Sent Events from Next.js API routes
B) WebSocket via separate service
C) Polling every 5 seconds

Context: 500 concurrent users, updates every 30 seconds on average, deployed on Vercel.

What are the tradeoffs? Recommend one with reasoning.

Use this for decisions, not implementation. Get the answer, THEN switch to Task Brief for building.

3. Context Management — The #1 Productivity Lever

Context Is Milk — It Spoils

When to Start Fresh (/new)

• Switching to unrelated task
• Context above 70%
• Claude starts repeating itself or making mistakes it didn't make earlier
• After shipping a feature (clean slate for next one)

When to Compact (/compact)

• Mid-task but context bloating from exploration
• After a debugging session (lots of error output consumed context)
• Before a complex implementation step

Context-Efficient Habits

1. Don't paste entire files — reference by path. Claude can read them.
2. Don't re-explain — if it's in CLAUDE.md, don't repeat it in prompts.
3. Use specific file paths — "Look at src/lib/auth.ts line 45" not "look at the auth code."
4. Close tangents — if Claude goes down a rabbit hole, redirect immediately. Don't let bad output consume context.
5. One concern per message — "Fix the auth bug AND refactor the database layer AND add tests" = context explosion. Sequential > parallel in a single session.

4. Sub-Agent Orchestration — Parallel Productivity

When to Use Sub-Agents

Task Tool Pattern (Claude Code native)

Use the Task tool to:
1. Research the Stripe API for subscription billing
2. Return: webhook event types we need, API calls for create/update/cancel, error codes to handle

Task tool spawns a sub-agent with its own context. Results come back summarized. Perfect for research that would bloat your main context.

Handoff Documents

When a sub-agent finishes complex work, have it write a HANDOFF.md:

## What Was Done
- Implemented Stripe webhook handler at src/app/api/webhooks/stripe/route.ts
- Added 4 event handlers: checkout.session.completed, invoice.paid, invoice.payment_failed, customer.subscription.deleted

## Key Decisions
- Used Stripe SDK v14 (not raw HTTP) for type safety
- Webhook signature verification via stripe.webhooks.constructEvent()
- Idempotency: check processed_events table before handling

## What's Next
- Wire up subscription status updates to user table
- Add retry logic for failed database writes
- E2E test with Stripe CLI: `stripe trigger checkout.session.completed`

## Gotchas
- Must use raw body (not parsed JSON) for signature verification
- Next.js App Router: export const runtime = 'nodejs' (not edge)

5. Debugging Workflow — Systematic, Not Random

The DEBUG Protocol

Describe the symptom (what you see vs. what you expect) Error output (paste full stack trace, not summary) Bisect (when did it last work? what changed?) Unit isolate (can you reproduce in a test?) Generate hypothesis (ask Claude for 3 possible causes, ranked)

Effective Bug Prompts

Bug: Users see stale data after updating their profile.

Expected: After PUT /api/profile, the profile page shows updated data.
Actual: Old data persists until hard refresh (Cmd+Shift+R).

Stack:
- Next.js 15 App Router
- Server component fetches user data
- Client component with form calls server action
- Server action calls db.update()

Hypothesis: Next.js is caching the server component fetch. Need to revalidate.

Can you confirm and show me the fix?

When Claude Gets Stuck

1. Add constraints — "The fix must not change the API contract" narrows the search space.
2. Share what you've tried — "I already tried revalidatePath('/profile') and it didn't work because..."
3. Ask for alternatives — "Give me 3 different approaches to solve this, with tradeoffs."
4. Fresh session — sometimes the context is poisoned. Start clean with just the bug description.

6. Refactoring Patterns — Safe Large-Scale Changes

The Refactoring Safety Net

Before any refactoring session:

Before we start refactoring:
1. Run the test suite and confirm it passes: `pnpm test`
2. Commit current state: `git add -A && git commit -m "chore: pre-refactor checkpoint"`
3. Create a branch: `git checkout -b refactor/[description]`

Safe Refactoring Prompts

Extract function:

Extract the email validation logic from src/lib/users.ts (lines 34-67) into a separate
function `validateEmail` in src/lib/validation.ts. Update all imports. Run tests after.

Rename across codebase:

Rename the `getUserData` function to `fetchUserProfile` across the entire codebase.
This includes: function definition, all call sites, all imports, all test references.
Run `pnpm test` and `pnpm lint` after to verify nothing broke.

Split large file:

src/lib/api.ts is 800 lines. Split it into:
- src/lib/api/users.ts (user-related functions)
- src/lib/api/projects.ts (project-related functions)
- src/lib/api/shared.ts (shared types and helpers)
- src/lib/api/index.ts (re-exports for backward compatibility)

Preserve all existing exports from the index file so no external imports break.
Run tests after each file move.

Multi-File Refactoring Rules

1. One type of change at a time — don't rename AND restructure AND optimize simultaneously.
2. Run tests after each step — catch breaks early, not after 20 files changed.
3. Preserve exports — use index.ts re-exports so consumers don't break.
4. Commit incrementally — one commit per logical change, not one giant commit.

7. Test-Driven Development With Claude Code

The TDD Loop

Step 1: "Write a failing test for: creating a user with valid email stores them in the database"
Step 2: [verify test fails for the right reason]
Step 3: "Now write the minimum code to make this test pass"
Step 4: [verify test passes]
Step 5: "Refactor the implementation — the test must still pass"

Why TDD Works Especially Well With Claude

• Tests are acceptance criteria — Claude knows exactly what "done" means.
• Immediate verification — no ambiguity about whether the code works.
• Prevents gold-plating — "minimum code to pass" stops Claude from over-engineering.
• Regression safety — future changes can't silently break working features.

Test-First Prompts

Write tests for a `calculateShipping` function that:
- Free shipping for orders over $100
- $5.99 flat rate for orders $50-$100
- $9.99 flat rate for orders under $50
- International orders: add $15 surcharge
- Express: 2x the base rate
- Edge cases: $0 order, negative amount (throw), exactly $50, exactly $100

Use vitest. Don't implement the function yet — just the tests.

Then: "Now implement calculateShipping to pass all tests."

8. Git Workflow With Claude Code

Pre-Work Checklist

git status          # Clean working tree?
git pull            # Up to date?
git checkout -b feat/[description]   # New branch

Commit Strategy

Claude Code Git Prompts

# After completing work:
"Commit the changes with an appropriate conventional commit message. 
Group related files into logical commits if there are multiple concerns."

# For PR creation:
"Create a PR description for these changes. Include:
- What changed and why
- How to test it
- Any migration steps needed
- Screenshots if UI changed"

9. Code Review Mode — Claude as Reviewer

Review Prompt Template

Review this code for:
1. Correctness — does it do what it claims?
2. Security — any injection, auth bypass, data leak risks?
3. Performance — N+1 queries, unnecessary re-renders, missing indexes?
4. Maintainability — clear naming, reasonable complexity, documented edge cases?
5. Testing — are the tests sufficient? Any missing cases?

Be specific. For each issue, cite the file:line and suggest a fix.
Skip style nits unless they affect readability.

Self-Review Before PR

I'm about to open a PR. Review all changed files (`git diff main`) for:
- Any hardcoded secrets or credentials
- TODO/FIXME/HACK comments that should be resolved
- Console.logs that should be removed
- Missing error handling
- Type assertions (as any) that should be proper types
- Missing tests for new public functions

10. Production Shipping Checklist

Before deploying any Claude-generated code:

P0 — Must Do

• [ ] All tests pass (pnpm test && pnpm test:e2e)
• [ ] Linting clean (pnpm lint)
• [ ] Type checking clean (tsc --noEmit)
• [ ] No hardcoded secrets in code (grep for API keys, tokens, passwords)
• [ ] Error handling exists for all external calls (DB, API, file I/O)
• [ ] Input validation on all user-facing endpoints
• [ ] Database migrations reviewed (no data loss, backward compatible)

P1 — Should Do

• [ ] Performance: no N+1 queries, no unbounded lists, pagination exists
• [ ] Logging: structured logs at appropriate levels (not console.log)
• [ ] Auth: all new endpoints have proper authorization checks
• [ ] Rate limiting on public endpoints
• [ ] Rollback plan documented (how to revert if broken)

P2 — Nice to Have

• [ ] Load test with expected traffic
• [ ] Monitoring alerts for new endpoints
• [ ] Documentation updated (API docs, README, CHANGELOG)
• [ ] Accessibility checked (if UI changes)

11. Anti-Patterns — What Kills Productivity

12. Speed Multipliers — Advanced Techniques

Slash Commands Reference

Model Selection Strategy

Keyboard Shortcuts

• Esc → interrupt Claude (stop generation if going wrong direction)
• Up arrow → edit last prompt (fix typo without re-typing)
• Tab → accept file edit suggestion

Session Stacking

For maximum throughput on a large feature:

1. Terminal 1: main implementation (Claude Code)
2. Terminal 2: tests for what Terminal 1 builds (Claude Code with /new)
3. Terminal 3: manual testing / running the app

13. Workflow Templates

New Feature Workflow

1. "Create branch feat/[name]"
2. "Write failing tests for [feature spec]"
3. "Implement minimum code to pass tests"
4. "Refactor — tests must stay green"
5. "Run full test suite + lint + typecheck"
6. "Commit with conventional commit message"
7. "Self-review: check diff for security, performance, missing edge cases"
8. "Create PR with description"

Bug Fix Workflow

1. "Create branch fix/[description]"
2. "Write a test that reproduces this bug: [paste error]"
3. "Fix the bug — test must pass"
4. "Run full test suite to check for regressions"
5. "Commit: fix: [description of what was wrong]"

Refactoring Workflow

1. "Create branch refactor/[description]"
2. "Run tests — confirm all green"
3. "Commit pre-refactor state"
4. "[Specific refactoring instruction]"
5. "Run tests — must still be green"
6. "Commit this step"
7. [Repeat 4-6 for each refactoring step]

Spike / Research Workflow

1. "Use Task tool to research [topic]. Return: key findings, API surface, gotchas, recommended approach."
2. [Read Task output]
3. "Based on the research, build a minimal proof of concept in /tmp/spike-[name]/"
4. [Evaluate POC]
5. "Spike looks good. Now implement properly in the main codebase."

14. Measuring Your Productivity

Track these weekly:

Natural Language Commands

*Built by AfrexAI — AI-native business tools that ship.*