Method

Research first

Before building, my agents read how it is really done: source code, documentation, measurements. I judge the results, and decisions are written down with their reason.

Written rules, enforced by tools

Agents and scripts work inside rules: plan first, an explicit go-ahead for anything outside the project, and checks that run as code rather than as promises.

Audit and counter-audit

Results are checked by a second, independent pass. In the end I still look myself, because scanners have missed what the eye caught.

Working with AI

I use Claude Code for almost all of the planning, code and execution. I set the goals, conventions and quality verdicts, and I test the results. Deleting originals, publishing and sending stay with me.

Context design. Every agent has a written role file: what it is responsible for, what it may touch, its goal and its process. So I rarely prompt from scratch.

Reverse prompting and a plan first. Agents ask clarifying questions and lay out their plan step by step before they act, a chain of thought I can correct early.

Grounding against hallucination. A large language model knows what a fact sounds like, not whether it is true. Facts therefore come from sources, measurements or the project index, and claims are checked before they go public.

Non-deterministic by nature. The same request can give different results. That is why every result gets an audit and an independent counter-audit.

Stewardship. I own the direction, the constraints and the final verdict; the agents do the typing.

Earlier versions of my Jellyfin scripts were made with ChatGPT, later ones with Claude. Unless a page says otherwise, scripts and plugins are tested on Windows 11 with Chrome, on Jellyfin Web 10.10.7 and 12.0+.

What did not work stays visible

Attempts that did not work stay in the articles. That is the point of keeping a record.