Sweet spot

• CRUD apps — AI nails forms, tables, auth
• Dashboards — data display, charts, filters
• Landing pages — layout, copy, responsive
• Internal tools — admin panels, data pipelines
• Browser extensions — small scope, clear API
• CLI tools — argument parsing, file I/O

Danger zone

• Real-time systems — WebSockets, concurrency bugs
• Payment processing — security-critical, edge cases
• Complex state — multi-step wizards, undo/redo
• Performance-critical — AI defaults to "works" not "fast"

Why this project?

• Small enough to build in 30 minutes
• Has real utility — you'll actually use it
• Touches all the phases: UI, data, API, deploy
• Easy to extend (AI summaries, shared shelves)

Stack

# The prompt we'll give Claude Code:

Build a bookmarking app called
"Link Shelf" with:

- Next.js + TypeScript
- SQLite via better-sqlite3
- Add link: URL + title + tags
- Search by title or tag
- Clean, minimal UI
- Dark/light mode toggle
- Deploy to Vercel

# Fix issues, add features one at a time

# "The tag filtering doesn't work when
#  I click a tag. Fix the handler."

# "Add a delete button with confirmation
#  dialog on each link card."

# "Make the search debounced - 300ms
#  delay before querying."

# "Add keyboard shortcut: Cmd+K to
#  focus the search bar."

# "Write tests for the bookmark CRUD
#  operations."

The iteration rules

• One feature per prompt — don't ask for 5 things at once
• Be specific about behavior — "debounced 300ms" not "make search better"
• Review every change — read what Claude wrote before moving on
• Test as you go — don't pile up 10 features then test
• Commit often — /commit after each working feature

# Deploy to Vercel (the easy path)
npm i -g vercel
vercel
# Follow prompts → live URL in 60 seconds

# Or Cloudflare Pages (free, edge)
npx wrangler pages deploy .next

# Or Railway (if you need a real DB)
railway init && railway up

Platform Cheat Sheet

High leverage

• Boilerplate — forms, CRUD, auth flows, API routes
• CSS & layout — responsive, animations, Tailwind classes
• Tests — unit tests, integration tests, fixtures
• Data transformations — parsing, mapping, validation
• Config files — tsconfig, ESLint, CI/CD, Docker
• Documentation — README, API docs, comments
• Migrations — framework upgrades, API changes

The CodeRabbit study (Dec 2025)

Common failure modes

• "Works on demo" — happy path works, edge cases crash
• Over-engineering — AI adds abstractions nobody asked for
• Stale patterns — uses outdated APIs or deprecated approaches
• Copy-paste debt — duplicates code instead of reusing
• Security blindness — 40-45% of AI code has vulnerabilities (Veracode)

The contradictory data

• GitHub Copilot study: 55% faster for simple, isolated tasks
• METR study (Jul 2025): experienced devs on real codebases were actually 19% slower with AI — despite believing they were 20% faster
• Biggest gains in greenfield code and unfamiliar languages
• Biggest losses in complex, existing codebases

The honest multiplier

Before you build

• Write the spec first — 30 min of spec saves hours of iteration
• Pick a boring stack — AI knows React/Next.js/Postgres better than your obscure framework
• Set up CLAUDE.md — teach the AI your conventions once
• Define "done" — ship the MVP, not the dream

While you build

• Commit after every working feature — you need rollback points
• Read every line AI writes — especially auth, payments, DB queries
• Use /model haiku for quick fixes — save Opus for architecture
• Write tests early — AI is great at tests, and they catch AI bugs
• Deploy early — find production issues before you have 50 features