Building & Delivering

Insights:

• Cursor-Small → good for small, scoped tasks (bug in one file, single component).
• GPT-5 → better for creative coding tasks (when requirements are vague).
• Claude-Sonnet-4 → better for precise, production-grade code (backend logic, multi-file changes).
• Always run npm run lint + npm run build after big tasks.
• Use build results as a heuristic of task quality.
• For 3rd-party services/APIs, first encapsulate them in a dedicated service module; verify it manually before broader use.
• Have AI generate comprehensive unit tests for that service until key paths and errors are covered.
• Immediately add a Cursor repo rule: these service modules may not be edited without your explicit permission.
• Avoid over-constraining UI implementation details; TailwindCSS guidance tends to work best.
• Smaller models may err initially but iterate and fix quickly with guidance—leverage for speed.

Actions:

• Let AI implement user story → stop after each phase → review.
• Extract reusable components later once UI stabilizes.
• Watch for repeated code, untranslated strings, unused components.
• For each 3rd-party integration: scaffold a service module and manually test basic calls.
• Ask AI to write unit tests for the service until coverage is satisfactory across success and failure paths.
• Lock the service via repo/cursor rules to disallow edits without your approval.