AI in practice¶

I have been using AI tooling seriously for five years. Not experimenting — working. The distinction matters.

What changed¶

For most of my career, the bottleneck in documentation was time. I knew what needed to be written. I had the structure in my head. Producing the actual text at the quality I wanted took longer than the project allowed. So documentation got cut, or compressed, or delegated to someone who did not understand the system well enough to write it accurately.

AI removed that bottleneck. Not by writing for me — by removing the friction between thinking and text. I provide the architecture, the source material, the editorial judgment. The AI provides velocity.

The result, in concrete terms: the libdrone documentation stack — master specification, variables file, payload SDK, hardware document, FreeCAD cookbook, shopping list, build guides — produced by one person, nights and weekends, 300 EUR in API costs. At a standard I would not have reached alone in the time available.

How I actually use it¶

Architecture first. I design the document structure before I write a word. What are the sections, what does each section answer, what is the single source of truth for each value. The AI does not design this. I do.

Source material in, structured text out. I feed technical facts — measurements, component specs, process steps — and get back prose that I then edit. The editing is significant. AI text without editorial judgment is recognisable and unconvincing.

Consistency enforcement. A 40-document stack with hundreds of cross-references is an editorial problem. AI is very good at checking consistency, finding contradictions, and flagging where a value appears in two places with different numbers.

The variables file discipline. Every numeric value in the libdrone documentation lives in one file. Every other document references the variable name, not the value. When something changes, you change one file. This is not an AI technique — it is an editorial technique I brought from the magazine world and applied to technical documentation.

What it does not do¶

AI does not know when something is wrong. It produces plausible text. Plausible is not the same as accurate. Every factual claim in every document I produce is verified against a primary source — the hardware, the measurement, the datasheet. The AI writes fast. I check everything.

AI does not have taste. It does not know that a sentence is technically correct but will confuse the reader who arrives without context. That judgment is editorial. It is what I bring.

Where I stand on it¶

The people who are nervous about AI in their field are mostly nervous about the wrong thing. It does not replace editorial judgment. It does not replace domain understanding. It does not replace the ability to design a documentation architecture that actually serves the person who needs it.

What it replaces is the excuse that good documentation takes too long. It does not, anymore. Which means the only remaining reason for bad documentation is that nobody cared enough to design it properly in the first place.