AI-assisted documentation¶
When I started building libdrone, I knew nothing about drones.
Not the aerodynamics. Not the components. Not the regulatory landscape, the hardware architecture, the flight controller ecosystem, or the physics of a 330mm hexaframe under load. I was a documentation specialist and a project manager who had decided to understand a complex technical domain from the inside.
The libdrone knowledge corpus — 160 articles, schema-governed, graph-connected — was built by that person. That is the actual story. Not a drone expert who documented what he knew. A documentation specialist who used AI to enter a domain he didn't, and built something that reads like it was written by an expert.
What AI actually unlocked¶
The standard assumption is that AI helps experts produce text faster. That is true, but it is the smaller half of the story.
The larger half: AI lets a person with deep structural and editorial skills operate in a domain they do not yet own.
I brought the architecture. I brought the editorial discipline — twenty years of it, including running a technology magazine and building enterprise documentation systems that actually got maintained. I brought the judgment about what a competent stranger needs in order to understand something without asking a question.
AI brought the domain. Not by inventing facts — by surfacing them, structuring them, connecting them, and flagging when something I had written contradicted something I had written three articles earlier.
The result is a corpus I could not have built alone. Not because I lacked the time. Because I lacked the knowledge — and AI closed that gap while I closed the structural one.
How the corpus is built¶
Every article covers exactly one concept. Fixed six-section structure: summary, concept, rationale, implementation, limitations, connections. Closed tag vocabulary. Explicit graph edges between articles.
The corpus can be walked. A build guide is not a separate document — it is a path through the graph, assembled from atoms, each maintained independently. When a component changes, one article updates. Everything that references it stays correct.
Schema design, tag vocabulary, connection logic — I made those decisions. The AI did not. Once the architecture existed, AI produced first drafts from source material: measurements, datasheets, build observations, design decisions. I edited. Sometimes lightly, sometimes heavily. The editorial judgment was always mine.
The AI was also the corpus memory: finding contradictions, identifying gaps, checking that the graph remained consistent as the corpus grew past what any single person can hold in working memory.
What it cost¶
One person. Approximately €150 in direct API costs over the project lifetime.
The standard is not "good enough for an open-source project." The standard is: could this be the reference documentation for an entire field? Documentation that outlasts any individual contributor, serves readers across multiple personas — student, builder, researcher, payload developer, institutional buyer — and remains coherent as the platform evolves.
That is the ambition. The €150 is the footnote.
The principle¶
AI did not make this possible by being intelligent. It made it possible by being the right tool in the right architecture.
The intelligence is the structure: one fact, one home, explicit connections, closed vocabularies, schema enforcement. These are editorial disciplines, not AI features.
What changes with AI is the entry cost to a new domain. For a person who already understands how knowledge should be organised, that entry cost is now very low.
That is worth understanding — especially in organisations where the distance between the people who understand AI and the people who understand the domain is the main obstacle to getting anything built.
→ libdrone.eu — the corpus in production
→ Documentation philosophy
→ Single source of truth
→ Work with me