What the process depends on
Research before tools. Most problems are half-solved in a vendor’s documentation, a licence page or a forum thread. I read those first, write down what I found and where, and only then choose a tool. The middleware case shows this with a dated evidence ledger.
A scope you can hold me to. The scope lists the deliverables, what I need from you, the timeline, the price and what counts as done. Changes go through the scope, not around it.
Checks before review, review before you. I test what can be tested automatically: links, forms, layouts at phone and laptop widths, packages and data. A second reviewer then reads the change line by line. You see the result after both.
Handover is a deliverable. Documentation written for the next person, a recorded walkthrough, and every account and file you control.
How I use AI
I started in September 2024 with MCP tool servers, using Serena for AI-assisted coding, before tools like Claude Code were widely available. Today I use AI assistants (Claude, ChatGPT/Codex, DeepSeek) every day for research, drafting, coding and review. The review process is documented. They make me faster, but they do not make decisions, and I remain accountable for the output. Before anything reaches a client, I check claims against their sources, test code and configuration on my own lab first, and treat anything I cannot verify as unconfirmed.
AI at work needs a careful setup and clear limits on what each tool can see and do. I can walk through how I set that up in an interview.
What I will not do
- Promise savings I have not measured.
- Copy a client’s material into another client’s work.
- Leave a system running that nobody on your side can explain.