AI Coding Toolkit

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: Dimension1 (Beginner)5 (Expert)Prompt quality"Fix this bug"Structured context + constraints + examplesContext managementPaste entire filesCurated context windows, .cursorrules, AGENTS.mdWorkflow integrationAd-hoc usageSystematic agent-first developmentOutput verificationAccept everythingReview, test, iterate before committingTool selectionOne tool for everythingRight tool for right task 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

Decision Guide: Which AI Coding Tool When?

ToolBest ForContext WindowAutonomy LevelCostGitHub CopilotLine/function completion, inline suggestionsCurrent file + neighborsLow (autocomplete)$10-19/moCursorFull-file editing, multi-file refactors, chatProject-aware (indexing)Medium (tab/chat/composer)$20/moWindsurf (Cascade)Autonomous multi-step tasks, flowsProject-aware + flowsHigh (agentic flows)$15/moClineVS Code extension, model-agnostic, transparentManual context + autoHigh (tool use, browser)API costsAiderTerminal-based, git-native, pair programmingRepo map + selected filesMedium-High (git commits)API costsClaude CodeCLI agent, complex multi-file tasksWorkspace-awareHigh (full agent)API costsOpenClawPersistent agent, cron, multi-surfaceWorkspace + memory + toolsVery High (autonomous)API costs

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: GitHub Copilot (always-on autocomplete) Cursor OR Windsurf (primary IDE) Claude Code OR Aider (terminal agent for complex tasks) Team: GitHub Copilot (org-wide) Cursor (primary IDE, .cursorrules in repo) 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)

System instructions (.cursorrules, AGENTS.md, CLAUDE.md, .windsurfrules) Explicit context (files you @mention or add to chat) Implicit context (open tabs, recent edits, project index) 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: Summarize, don't dump — Instead of pasting a 500-line file, describe what it does and paste only the relevant section Use @mentions — @file.ts instead of copy-paste (tool-specific) Create reference docs — One-page architecture summaries the AI can reference Prune conversation — Start new chats for new tasks; stale context = hallucinations 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.

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]

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

Cursor

FeaturePower MoveTab completionLet it complete 3-5 tokens before accepting — catches wrong predictions earlyCmd+K (inline edit)Select ONLY the exact lines to change — less context = more accurateChat@file to add context, @codebase for project-wide questionsComposerMulti-file changes — describe the full feature, let it edit across files.cursorrulesProject-specific AI instructions — commit to repo for team alignmentNotepadsReusable context (API docs, design docs) — attach to any chat 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)

FeaturePower MoveCascade flowsMulti-step autonomous tasks — it can read, write, run terminalWrite modeDirect file editing with AIChat modeDiscussion without editing.windsurfrulesProject context fileTurbo modeFaster, less accurate — good for simple tasks 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

FeaturePower MoveModel selectionSwitch models per task (cheap for simple, expensive for complex)Tool useReads files, runs commands, opens browser — full agentTransparencyShows every action before executing — audit everythingCustom instructionsPer-project system promptsAuto-approveConfigure which actions need approval 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

FeaturePower Move/add filesExplicitly control which files the AI can see/edit/read filesRead-only context (reference files)/architectTwo-model approach — architect plans, editor implementsRepo mapAuto-generates codebase summary for contextGit integrationEvery change is a commit — easy rollback 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

FeaturePower MoveFull agentReads files, writes code, runs tests, git operationsCLAUDE.mdProject instructions file — auto-loadedSub-agentsSpawn parallel workers for complex tasksMemoryPersistent across sessions (project-level) 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

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

FailureDetectionFixHallucinated APICode uses functions that don't existCheck library docs before acceptingOutdated patternsUses deprecated APIs (React class components)Specify versions in contextMissing error handlingHappy path only, no try/catchAsk specifically for error casesSecurity holesInline secrets, missing auth, XSSSecurity review as separate stepOver-engineering5 files for a 20-line solutionAsk for simplest possible solutionWrong abstractionsPremature generalizationSpecify "don't abstract, keep concrete"Test theaterTests that pass but test nothingReview test assertions specificallyCopy-paste bugsDuplicated logic with subtle differencesCheck for patterns, extract helpers

The 3-Read Review

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

Token Cost Awareness

ModelInput $/1M tokensOutput $/1M tokensBest ForGPT-4o mini$0.15$0.60Simple completions, formattingClaude Haiku$0.25$1.25Quick edits, simple questionsGPT-4o$2.50$10.00Complex code generationClaude Sonnet$3.00$15.00Complex code, long contextClaude Opus$15.00$75.00Architecture, hardest problemso3$10.00$40.00Complex reasoning, algorithms

Cost Reduction Strategies

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

Monthly Cost Benchmarks (Full-Time Developer)

Usage LevelEstimated Monthly CostLight (Copilot + occasional chat)$20-40Medium (Cursor Pro + daily chat)$40-80Heavy (API-based agents, complex tasks)$80-200Power user (autonomous agents, all day)$200-500+

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

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

Anti-PatternWhy It FailsDo This InsteadPrompt and prayNo verification = bugs in productionAlways review, always testPaste the whole codebaseOverwhelms context, increases costCurate relevant files onlyNever start new chatsStale context → hallucinationsNew task = new chatTrust without readingAI generates plausible but wrong codeRead every lineSkip tests because AI wrote itAI code has bugs tooTest AI code MORE, not lessUse one model for everythingWaste money on simple tasksTier models by complexityNo project rules fileAI guesses your conventionsWrite .cursorrules / CLAUDE.mdVague promptsGarbage in, garbage outUse SPEC frameworkOver-relianceSkill atrophy, can't debug AI outputUnderstand what AI generatesIgnoring securityAI doesn't prioritize securityExplicit security review step

AI-Assisted Development Quality Score (0-100)

DimensionWeightCriteriaContext engineering20%Rules files, curated context, fresh chatsPrompt quality15%SPEC framework, task-appropriate templatesVerification rigor20%Review checklist, test coverage, security reviewTool selection10%Right tool for task, model routingCost efficiency10%Tiered usage, context management, batch tasksOutput quality15%Code correctness, maintainability, no driftWorkflow integration10%Systematic process, team alignment

Weekly Self-Review Questions

What was my best AI-assisted output this week? What made it good? Where did AI waste my time? What went wrong with context/prompts? Am I reviewing thoroughly enough, or rubber-stamping? What prompt patterns worked well? Add to prompt library. 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

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