Why does teaching an AI how to do something often beat handing it more facts about the subject?

This explores why giving an AI the how of a task (steps, rules of thumb, design principles) tends to improve what it produces more than giving it more facts about the subject. The corpus points to one main reason: facts and procedures behave differently. Facts are narrow and have to be stored, while procedures are broad and transfer. An analysis of 5 million pretraining documents found that reasoning draws on how-to material spread across many diverse sources, while factual recall depends on memorizing specific facts from specific documents (Does procedural knowledge drive reasoning more than factual retrieval?). When researchers looked at what agent skills actually do, about 66% of successful cases worked by anchoring the agent's actions, and under 5% worked by supplying a missing fact (Do skills teach procedures or inject missing facts?).

Authoring tasks sharpen this because there is rarely a fact to look up. The quality of a slide deck, a design or a program comes from the order of steps and the judgment applied at each one. That judgment is often the tacit knowledge specialists carry in their heads. In one industrial case, writing a specialist's rules and design principles into an agent's scaffolding raised output quality by 206%, and non-experts using it were rated at expert level. The gain came from externalizing that expertise, not from a bigger model (Can codified expertise let non-experts match specialist output?). Distilling tutorial videos, code repositories and articles into a reusable skill library lifted artifact quality by about 12 points across seven authoring domains (Can tutorial videos teach software agents reusable skills?).

Procedures also have practical advantages over facts. They can be mined from source code with no prior agent experience and checked by rebuilding the code (Can source code replace experience as skill raw material?). They compound: routines extracted from past tasks and stacked hierarchically gained more as the new tasks looked less like the old ones (Can agents learn reusable sub-task routines from past experience?). A skill document can even be trained like model weights, with edits accepted only if they improve results on held-out tests. That adds no inference cost and carries over to other models (Can skill documents be optimized like neural network weights?). Putting the procedure in explicit code also lets each step see only the context it needs (Can algorithms control LLM reasoning better than LLMs alone?).

The corpus also draws limits, so procedure is not simply better than facts. Prompting can only reorganize what the model already knows, and no prompt can supply knowledge missing from training (Can prompt optimization teach models knowledge they lack?). Knowledge and reasoning seem to sit in different layers of the network, so training for reasoning can help math and hurt medicine (Why does reasoning training help math but hurt medical tasks?). Procedures are also fragile in their own ways. Skills fail when retrieved wrongly, invoked out of context or followed too rigidly (Do skills teach procedures or inject missing facts?). Playbooks erode when rewritten wholesale, which is why incremental updates matter (Can context playbooks prevent knowledge loss during iteration?). One gap remains: no note here runs a controlled facts-versus-procedure comparison on authoring specifically. The case is assembled from studies of skills, scaffolding and pretraining data that point the same way.