Alberto Fortin’s recent post on dialing back LLM usage resonated with me. Like many developers, I’ve found AI to be both a productivity godsend and a source of friction. It can feel like magic — until it quietly leads you down the wrong path.
Here’s my take (as of today, it may all change tomorrow): AI can be a significant coefficient of productivity, but it has to be used wisely.
Don’t Trust It Blindly
AI doesn’t understand your architecture, your domain, or your standards (even if you use MCPs, RAGs, etc). Even when it produces elegant-looking code, there’s often a snippet hidden inside that you’d never use in production. That’s the paradox: it can look solid while introducing subtle bugs or anti-patterns.
So here’s rule number one: you are the architect.
Use LLMs to brainstorm ideas, explore alternate approaches, or ask for opinions on an implementation plan. You might receive a suggestion that hadn’t occurred to you — and that’s useful. But don’t confuse novelty with correctness. The architectural decisions must remain yours.
Always Review the Code — Every Line
A good LLM output can save you time. But you must read through the results. Every. Single. Line.
Ask yourself: if I have to review everything anyway, is it even worth using?
In most cases, yes — because you’re skipping boilerplate, scaffolding, and tedious setups. That time saved is real. But what you’re not skipping is the thinking.
Craftsmanship takes time. That hasn’t changed.
Use AI Where It Shines
Here’s where AI excels:
- Rewriting repetitive scaffolding
- Converting one style to another (e.g. JSON to YAML, SQL to ActiveRecord)
- Suggesting test coverage for existing code
- Generating documentation or summaries
Here’s where AI struggles:
- Making high-level design decisions
- Understanding business context
- Writing robust production-ready code from scratch
The Balanced Mindset
I still use LLMs every day. But I treat them like a junior dev with infinite patience and energy — helpful, fast, and occasionally wrong.
AI is a productivity multiplier, not a replacement for skill.
The moment you surrender judgment to it, the benefits start to erode. But if you remain thoughtful, critical, and in control, LLMs can absolutely make you a more efficient, expressive developer.
This post is a response to “Why I’m Dialing Back My LLM Usage” by Alberto Fortin on Zed. I agree with much of it — but believe the answer is not to retreat from AI, but to wield it with discipline.