afrexai-ai-coding-toolkit（afrexai AI 编码工具包）

AI Coding Toolkit — Master Every AI Coding Assistant

The complete methodology for 10X productivity with AI-assisted development. Covers Cursor, Windsurf, Cline, Aider, Claude Code, GitHub Copilot, and more — tool-agnostic principles that work everywhere.

Phase 1: Quick Assessment — Where Are You?

Rate yourself 1-5 on each:

Score interpretation:

• 5-10: Read everything — you'll 10X your output
• 11-18: Skip to Phase 4+ for advanced techniques
• 19-25: Focus on Phase 8-10 for mastery patterns

Phase 2: Tool Selection Matrix

Decision Guide: Which AI Coding Tool When?

Selection Decision Tree

Need autocomplete while typing?
  → GitHub Copilot (layer it with any other tool)

Working in VS Code/IDE?
  ├─ Want integrated editor experience? → Cursor or Windsurf
  ├─ Want model flexibility + transparency? → Cline
  └─ Want minimal config, just works? → Cursor

Working in terminal?
  ├─ Want git-native pair programming? → Aider
  ├─ Want full agent with tools? → Claude Code
  └─ Want persistent autonomous agent? → OpenClaw

Building complex multi-file features?
  → Cursor Composer or Windsurf Cascade or Claude Code

Need autonomous background work?
  → OpenClaw (cron, heartbeats, multi-session)

Recommended Stack (Layer These)

Solo developer:

1. GitHub Copilot (always-on autocomplete)
2. Cursor OR Windsurf (primary IDE)
3. Claude Code OR Aider (terminal agent for complex tasks)

Team:

1. GitHub Copilot (org-wide)
2. Cursor (primary IDE, .cursorrules in repo)
3. CI/CD AI review (automated PR review)

Phase 3: Context Engineering — The #1 Skill

Context is everything. The quality of AI output is directly proportional to the quality of context you provide.

The Context Hierarchy (Most → Least Important)

1. System instructions (.cursorrules, AGENTS.md, CLAUDE.md, .windsurfrules)
2. Explicit context (files you @mention or add to chat)
3. Implicit context (open tabs, recent edits, project index)
4. Model knowledge (training data — least reliable for your codebase)

Project Rules File Template

Create at project root. Name depends on tool:

• Cursor: .cursorrules
• Windsurf: .windsurfrules
• Claude Code: CLAUDE.md
• Aider: .aider.conf.yml + convention docs
• OpenClaw: AGENTS.md

# [PROJECT] — AI Coding Context

## Project Overview
- Name: [project name]
- Stack: [e.g., Next.js 14 + TypeScript + Tailwind + Drizzle + PostgreSQL]
- Architecture: [e.g., App Router, server components by default]
- Monorepo: [yes/no, structure if yes]

## Code Standards (ENFORCE STRICTLY)
- TypeScript strict mode (`tsc --noEmit --strict`)
- Max 50 lines per function, 300 lines per file
- One responsibility per file
- Naming: camelCase functions, PascalCase types, SCREAMING_SNAKE constants
- Imports: named imports, no default exports
- Error handling: explicit try/catch, typed errors, no silent catches

## Patterns to Follow
- [Pattern 1 with example]
- [Pattern 2 with example]
- [Pattern 3 with example]

## Anti-Patterns (NEVER DO)
- [Anti-pattern 1]
- [Anti-pattern 2]
- [Anti-pattern 3]

## File Structure

src/ components/ # React components lib/ # Shared utilities server/ # Server-only code db/ # Database schema + queries types/ # Shared TypeScript types

## Testing
- Framework: [vitest/jest/pytest]
- Pattern: AAA (Arrange, Act, Assert)
- Naming: `should [expected behavior] when [condition]`
- Coverage target: [80%+]

## Dependencies
- Approved: [list]
- Banned: [list with reasons]

## Common Commands
- `npm run dev` — start dev server
- `npm run test` — run tests
- `npm run lint` — lint + typecheck
- `npm run build` — production build

Context Window Management

The 80/20 Rule: 80% of your context should be the specific files/functions relevant to the task. 20% is project conventions and standards.

Context Compression Techniques:

1. Summarize, don't dump — Instead of pasting a 500-line file, describe what it does and paste only the relevant section
2. Use @mentions — @file.ts instead of copy-paste (tool-specific)
3. Create reference docs — One-page architecture summaries the AI can reference
4. Prune conversation — Start new chats for new tasks; stale context = hallucinations
5. Tree command — Give the AI your project structure: tree -I node_modules -L 3

The Context Refresh Rule

Every 5-10 messages, check: Is the AI still tracking correctly? If it starts hallucinating file names, functions, or making wrong assumptions — start a new chat with fresh context. Context is milk. It spoils.

Phase 4: Prompt Engineering for Code

The SPEC Framework (Structure, Precision, Examples, Constraints)

Bad prompt:

Fix the login bug

Good prompt (SPEC):

## Structure
Fix the authentication flow in `src/auth/login.ts`

## Precision
- The login function throws "user not found" even when the user exists
- Error occurs on line 42 when querying by email (case-sensitive match)
- PostgreSQL query uses exact match but emails are stored lowercase

## Examples
- Input: "User@Example.com" → should match "user@example.com" in DB
- Current behavior: returns null
- Expected: returns user record

## Constraints
- Don't change the database schema
- Use the existing `normalizeEmail()` utility from `src/utils/email.ts`
- Add a test case for case-insensitive lookup
- Keep the existing error handling pattern (throw AppError)

Prompt Templates by Task Type

Feature Implementation:

Implement [feature] in [file/location].

Requirements:
1. [Requirement with acceptance criteria]
2. [Requirement with acceptance criteria]
3. [Requirement with acceptance criteria]

Constraints:
- Follow existing patterns in [reference file]
- Use [specific library/approach]
- Include error handling for [edge cases]
- Write tests in [test file location]

Reference: Here's how similar feature [X] was implemented:
[paste relevant code snippet]

Bug Fix:

Bug: [description]
File: [path]
Steps to reproduce: [1, 2, 3]
Expected: [behavior]
Actual: [behavior]
Error: [paste error message/stack trace]

Fix constraints:
- Don't change [protected areas]
- Add regression test
- Explain root cause before fixing

Refactoring:

Refactor [file/module] to [goal].

Current state: [describe current architecture]
Target state: [describe desired architecture]
Motivation: [why — performance, readability, maintainability]

Rules:
- Preserve all existing behavior (no functional changes)
- Keep all existing tests passing
- Break into small, reviewable commits
- Each commit should be independently deployable

Code Review:

Review this code for:
1. Correctness — logic errors, edge cases, race conditions
2. Security — injection, auth bypass, data exposure
3. Performance — N+1 queries, unnecessary allocations, missing indexes
4. Maintainability — naming, complexity, test coverage

Be specific: quote the line, explain the issue, suggest the fix.
Skip style/formatting — linter handles that.

[paste code]

Phase 5: Workflow Patterns — Agent-First Development

Pattern 1: Test-Driven AI Development (TDD-AI)

1. Write the test first (yourself or with AI help)
2. Ask AI to implement the code that passes the test
3. Run tests — verify green
4. Ask AI to refactor while keeping tests green
5. Review the final code yourself

Why this works: Tests are specifications. The AI writes better code when it has a concrete target. You catch hallucinations immediately.

Pattern 2: Scaffold → Fill → Review

1. Ask AI to scaffold the architecture (file structure, interfaces, types)
2. Review and approve the scaffold
3. Ask AI to fill in implementation file by file
4. Review each file individually
5. Integration test the full feature

Why this works: You maintain architectural control. The AI handles the grunt work. Errors are caught at each layer.

Pattern 3: Conversation Threading

Chat 1: Architecture discussion → decisions documented
Chat 2: Implementation of Component A (reference architecture doc)
Chat 3: Implementation of Component B (reference architecture doc)
Chat 4: Integration + testing

Why this works: Fresh context per component prevents drift. Architecture doc provides continuity.

Pattern 4: AI Pair Programming (Aider/Claude Code)

1. Start session with repo context
2. Describe the task in natural language
3. AI proposes changes as git diffs
4. Review each diff before accepting
5. AI commits with meaningful messages
6. You handle edge cases and integration

Pattern 5: Autonomous Agent Workflow (OpenClaw/Claude Code)

1. Define task in structured format (acceptance criteria, constraints)
2. Agent plans → executes → verifies (reads files, runs tests)
3. Agent creates PR/branch with changes
4. You review the complete changeset
5. Iterate on feedback

Phase 6: Tool-Specific Power Moves

Cursor

Cursor Pro Tips:

• Use @git to reference recent changes
• Use @docs to reference official library documentation
• Create .cursor/rules/ directory for multiple rule files by domain
• "Apply" button to accept chat suggestions directly into code

Windsurf (Cascade)

Windsurf Pro Tips:

• Cascade excels at multi-file refactors — give it the full scope
• Use "undo flow" to revert entire multi-step changes
• Pin important files in context
• Let it read error output from terminal to self-fix

Cline

Cline Pro Tips:

• Set spending limits to prevent runaway API costs
• Use cheaper models (Haiku/GPT-4o-mini) for simple tasks
• Enable "diff mode" to see exact changes before applying
• Create task-specific instruction files

Aider

Aider Pro Tips:

• Use --architect flag for complex features (planner + implementer)
• /drop files you don't need to free context window
• --map-tokens to control repo map size
• Run aider --model claude-sonnet-4-20250514 for best code quality

Claude Code

Claude Code Pro Tips:

• Write a comprehensive CLAUDE.md — it's your biggest leverage
• Use "plan mode" first for complex tasks, then "implement"
• Let it run tests and self-correct — don't interrupt the loop
• Use /compact when context gets long

Phase 7: Code Quality Guardrails

The Trust-But-Verify Checklist

After every AI-generated change:

• [ ] Read every line — don't blindly accept. AI hallucinates plausible-looking code
• [ ] Check imports — AI often imports non-existent modules or wrong versions
• [ ] Verify function signatures — parameter names, types, return types
• [ ] Test edge cases — AI optimizes for the happy path
• [ ] Check for security — hardcoded secrets, missing auth checks, SQL injection
• [ ] Run the tests — if tests pass, good. If no tests exist, write them first
• [ ] Check for drift — did it change files you didn't ask it to change?
• [ ] Verify dependencies — did it add packages? Are they real? Are they secure?

Common AI Code Failures

The 3-Read Review

1. Skim read — Does the structure make sense? Right files, right approach?
2. Logic read — Does each function do what it claims? Edge cases handled?
3. Integration read — Does it work with the rest of the codebase? Breaking changes?

Phase 8: Cost Optimization

Token Cost Awareness

Cost Reduction Strategies

1. Tier your usage — Simple tasks → cheap model. Complex → expensive model
2. Reduce context — Every unnecessary file in context costs money
3. Start new chats — Long conversations accumulate expensive history
4. Use autocomplete for simple stuff — Copilot is flat-rate, much cheaper per completion
5. Cache project context — Use rules files instead of re-explaining every chat
6. Batch related tasks — Handle related changes in one conversation

Monthly Cost Benchmarks (Full-Time Developer)

Phase 9: Team Adoption

Rolling Out AI Coding Tools to a Team

Week 1-2: Foundation

• Choose primary tool (Cursor or Windsurf recommended for teams)
• Create .cursorrules / .windsurfrules committed to repo
• Run a 1-hour workshop: basics, prompt techniques, verification
• Set team guidelines (review requirements, security rules)

Week 3-4: Practice

• Daily 15-min "AI wins" standup share
• Pair sessions: experienced + new user
• Collect common prompts into team prompt library
• Monitor and address concerns (quality, dependency)

Month 2: Optimization

• Measure: time-to-PR, bugs-per-feature, developer satisfaction
• Iterate on .cursorrules based on team feedback
• Create task-specific prompt templates in shared docs
• Address skill gaps: who's using it well, who needs help?

Month 3: Systemization

• AI-assisted PR review as CI step
• Automated test generation for new features
• Custom slash commands / snippets for team workflows
• Quarterly review: ROI, quality metrics, tooling updates

Team Guidelines Template

# AI Coding Guidelines — [Team Name]

## Approved Tools
- [Tool 1] for [use case]
- [Tool 2] for [use case]

## Rules
1. AI-generated code gets the SAME review rigor as human code
2. Never paste proprietary/customer data into AI tools without approved data handling
3. All AI-generated tests must be reviewed for assertion quality
4. Security-sensitive code (auth, payments, PII) requires human-first approach
5. Commit messages should NOT mention AI — own the code you commit

## Quality Gates
- [ ] Typecheck passes (`tsc --noEmit --strict`)
- [ ] All tests pass
- [ ] No new warnings
- [ ] Manual review of all AI-generated code
- [ ] Security-sensitive areas reviewed by security champion

Phase 10: Advanced Patterns

Multi-Agent Architecture for Development

Task: Build feature X

Agent 1 (Architect): Plans the approach, defines interfaces
Agent 2 (Implementer): Writes the code
Agent 3 (Tester): Writes and runs tests
Agent 4 (Reviewer): Reviews for quality, security, patterns

Orchestrator: Coordinates, resolves conflicts, maintains context

Self-Healing Development Loop

1. Agent writes code
2. Agent runs tests
3. Tests fail → agent reads error, fixes code
4. Repeat until tests pass
5. Agent runs linter
6. Lint fails → agent fixes
7. All green → create PR

The Prompt Library Pattern

Maintain a prompts/ directory in your project:

prompts/
  feature-implementation.md
  bug-fix.md
  refactoring.md
  code-review.md
  test-generation.md
  migration.md
  documentation.md

Each file is a reusable prompt template. Reference them: "Follow the template in prompts/feature-implementation.md"

Model Routing Strategy

task_routing:
  autocomplete: copilot  # Always-on, flat rate
  simple_edit: haiku     # Quick, cheap
  feature_impl: sonnet   # Good balance
  architecture: opus     # When it matters
  debugging: sonnet      # Needs to reason about code
  documentation: haiku   # Simple transformation
  security_review: opus  # Can't afford mistakes
  test_generation: sonnet # Needs understanding of code logic

Phase 11: Anti-Patterns — What NOT to Do

Phase 12: Scoring & Continuous Improvement

AI-Assisted Development Quality Score (0-100)

Weekly Self-Review Questions

1. What was my best AI-assisted output this week? What made it good?
2. Where did AI waste my time? What went wrong with context/prompts?
3. Am I reviewing thoroughly enough, or rubber-stamping?
4. What prompt patterns worked well? Add to prompt library.
5. Am I over-relying on AI for things I should understand deeply?

Monthly Metrics

• Acceleration factor: Tasks completed per day vs pre-AI baseline
• Bug rate: Bugs in AI-assisted code vs manual code
• Cost per feature: API spend / features shipped
• Context efficiency: Average conversation length before drift
• Coverage: % of codebase with AI-assisted tests

Quick Reference: Natural Language Commands

1. "Set up AI coding for [project]" — Generate rules file + tool recommendations
2. "Write a prompt for [task type]" — Generate SPEC-formatted prompt template
3. "Review this AI output" — Run the Trust-But-Verify checklist
4. "Compare [tool A] vs [tool B] for [use case]" — Tool selection analysis
5. "Optimize my AI coding costs" — Analyze usage and suggest model routing
6. "Create a team AI coding guide" — Generate team guidelines document
7. "Debug why AI keeps [hallucinating X]" — Context diagnosis
8. "Set up test-driven AI workflow for [feature]" — TDD-AI pattern guide
9. "Create prompt library for [project type]" — Generate prompt templates
10. "Score my AI coding maturity" — Run the quality assessment
11. "Onboard [person] to AI coding" — Generate training plan
12. "Audit AI coding security practices" — Security review checklist