AI & Engineering 7 min read May 10, 2025
🤖

My Exact Claude Setup for Daily Engineering Work

1.2k views 890 unique 47

I've been using Claude daily for engineering work for over a year. I've tried every workflow, killed most of them, and kept what actually moves the needle. Here's the exact setup I use — no fluff, no affiliate links.

The philosophy first

Most people use AI like a smarter Stack Overflow. Type question, get answer, copy code. That's fine, but it's leaving 80% of the value on the table.

The shift that changed everything for me: treat the AI as a collaborator on a problem, not a dispenser of answers. That means sharing context, iterating, and sometimes arguing with it.

MCP servers I actually use

MCP (Model Context Protocol) lets Claude talk to external tools. I run three servers daily:

filesystem — Claude can read and write files in my project. This sounds simple and it is. But it means I can say 'refactor the auth module' and it can actually do it across multiple files.

github — Claude can read PRs, issues, and code. Useful for context when asking about existing patterns.

postgres — direct DB access for debugging data issues. Saved me hours on a migration bug last month.

My prompting workflow (SDD-lite)

For any non-trivial task I follow Spec-Driven Development:

  1. **Spec first** — I describe the feature or fix in plain English, including constraints and edge cases.
  2. **Claude drafts the spec** — I ask it to restate my requirements back as a structured spec.
  3. **I edit the spec** — This is where I catch misunderstandings.
  4. **Then we write code** — Against the agreed spec.

The spec step feels slow. It saves me 2x the time in debugging.

The hooks setup

I use Claude Code hooks to run tests automatically after edits. If tests fail, Claude sees the output and self-corrects. About 60% of the time it fixes the issue without me doing anything.

What doesn't work

Using it for architecture decisions without context. It gives confident-sounding answers based on common patterns. Your problem is probably not common.

Long sessions without checkpoints. Context drift is real. I checkpoint every hour-ish with a summary.

Trusting output without reading it. Still your code. Still your prod. You need to understand what you're shipping.

The honest take

AI hasn't made me a 10x engineer. It's made me a 1.6x engineer who spends way less time on boilerplate and way more time on the actual thinking. That's the real win.

All posts