AI in my workflow
AI writes the code. I stay in control.
I mainly build with Claude Code. But there is one absolute rule: I never let it code blindly. I’m right behind it, physically, watching every move.
My method, step by step
- 01
A real need
Everything starts from a concrete problem: something that bothers my family or friends, or an idea that answers a need I see around me. No project for the sake of it.
- 02
I think it through
Before touching anything, I work out how I could improve the situation. What would really be useful? What is needlessly complicated?
- 03
Plan mode
I talk it through with the AI in Plan mode. Together we define the stack, the system, the features, the type of database, where and how it’s accessed, and whether we need this or that. I give it everything I want.
- 04
Supervised development
Now the code starts, but I’m behind the screen. I watch what’s done, I ask questions, I refuse what I don’t understand or don’t find clean.
- 05
Security check
I first look at what’s sensitive: the files dealing with access, secrets and security. No API key or access code is ever to be hard-coded in the source.
My non-negotiable rules
- No plan, no code.
- No hard-coded secrets: no API key, password or access code in the files.
- I review sensitive files myself before moving on.
- I understand what’s written before accepting it.
- I stay present throughout development, never on autopilot.
- Security comes before speed.
Why this approach
Letting an AI write code without looking is fast, but it’s also how keys end up published and flaws slip through. My cybersecurity vocational diploma makes me especially aware of those risks.
I believe the coder’s job is evolving: AI writes fast, and a developer’s value lies in knowing what to build, how, and checking that it’s safe. That’s what I call being an Engineer Code.
This philosophy is also Elysium’s: a human-approved plan, agents with limited permissions, masked secrets.
An AI-assisted coder, not a spectator of the AI.