A "tutorial factory" driven by agents: 26 guides published in five months for 1.5 hours of human work per week

Researchers describe a training content production chain where agents write, humans validate, and public engagement rewrites the generation rules.

The technical tools that professionals learn evolve in a matter of weeks, and the guides that describe them age just as quickly. LLMs make writing a tutorial cheap, but most systems produce content once and for all, without ever checking what the audience does with it. One team presents an alternative that treats educational production as a manufacturing line punctuated by controls.

Scripts, agents, and human validation

The pipeline is a typed graph with four kinds of nodes: deterministic scripts, model steps evaluated on a fixed grid, more open agent phases, and human approval gates. It distinguishes between a "demand" layer, fed by videos and online communities to choose topics and tone, and an "authority" layer, built on official documentation, to which each published statement must be able to be attached. A feedback loop rewrites the generation rules from the engagement statistics of each publication.

In five months of continuous operation, the system published 26 documented guides for approximately 1.5 hours of operator time per week. The loop highlighted four measurable effects, which became mandatory rules for subsequent productions, including a hook style that performed 1.83 times better than the benchmark during its first analysis period.