Engineering · 2025-02-01

The AI tools I actually use to ship software faster

A practical rundown of the AI tools in my daily development workflow — what's genuinely useful, what's overhyped, and how I use them together.

The honest framing

There's a lot of noise around AI development tools. Everyone has opinions. Here's mine, filtered through actually shipping products with these tools, not just experimenting with them.

I'll cover what I use, how I use it, and where it falls short.

Claude Code

What it is: Anthropic's CLI agent for software development. It reads your codebase, makes edits, runs commands, and works through complex tasks.

How I use it: For the majority of implementation work — especially boilerplate, refactoring, debugging, and building out features once I know what I want to build. I run it in a terminal alongside my editor.

Where it shines:

Large refactors that would be tedious manually
Debugging with full codebase context (it can search across files, read related components)
Generating tests
First drafts of new components or modules that follow existing patterns in the codebase

Where it falls short:

Complex UI work where you need to see the result in real time — still need to iterate manually
Novel architectural decisions — it'll give you an answer, but the judgment call is still yours
Anything requiring real-world context it doesn't have (your specific business logic, your users' behavior)

My workflow: I describe what I want at a reasonably high level, review what it produces, course-correct, and repeat. It's a fast feedback loop when you treat it as a collaborator, not a vending machine.

Cursor

What it is: An AI-native code editor (VS Code fork) with context-aware autocomplete and an inline chat interface.

How I use it: For active coding sessions where I'm writing new code or making targeted edits. The Tab autocomplete is genuinely good — it predicts multi-line completions that match the surrounding code style.

The best feature: @codebase context in chat. You can ask questions about your entire codebase and get accurate answers. "Where is the booking state managed?" "What does this function do?" Faster than searching.

Vs. Claude Code: They're complementary. Cursor is for in-editor coding; Claude Code is for larger tasks run from the terminal. I use both.

Anthropic API (Claude directly)

What it is: Direct access to Claude models via API for building AI-powered features into products.

How I use it: For the LLM layer in products I build — the WhatsApp agent, intent classification, anything that requires language understanding.

Model choice: For production use cases, claude-sonnet-4-5 is the right balance of quality and cost. For complex reasoning or agentic tasks, claude-opus-4-5. For fast, cheap inference where quality matters less, claude-haiku-4-5.

The feature I use most: Tool use (function calling). For agentic systems where the LLM needs to take actions, defining a set of tools and letting Claude decide which to call is the right pattern. It's more reliable and debuggable than trying to parse free-form LLM output.

n8n

What it is: Open-source workflow automation (self-hosted alternative to Zapier/Make).

How I use it: For multi-step automation workflows that don't need custom code — connecting external APIs, scheduling jobs, triggering actions based on events.

Why not Zapier: n8n is free to self-host, has more flexibility, and you can run custom JavaScript nodes when you need logic that doesn't fit a pre-built integration.

Where it fits: n8n handles the orchestration; Claude handles the intelligence; custom code handles the business logic. They're different layers of the same system.

Supabase

What it is: Open-source Firebase alternative — Postgres database, auth, real-time, storage, edge functions.

How I use it: As the backend for almost everything I build. It handles auth, database, and real-time subscriptions out of the box.

AI-relevant use: The vector storage extension (pgvector) makes it easy to add semantic search and RAG (retrieval-augmented generation) to products without a separate vector database.

What I don't use (and why)

GitHub Copilot: Cursor's autocomplete is better, in my experience. I switched and didn't look back.

ChatGPT for coding: Claude is more accurate for code, maintains context better, and is less likely to confidently produce wrong answers. I use ChatGPT for non-coding research occasionally.

No-code AI builders: For anything I'd put in production, I want to understand and control the code. No-code tools are fast to start and slow to customize.

The actual productivity multiplier

The combination of Claude Code + Cursor + a well-chosen backend (Supabase) means a solo developer can now ship at the pace of a small team. Not because the AI is magical, but because:

1. The boilerplate cost of building software has dropped dramatically 2. Debugging cycles are shorter 3. You can maintain more context across the codebase

The bottleneck has shifted from "can I build this?" to "do I know what to build and why?" That's a better problem to have.