Assigning glycemic index (GI) values to food composition databases is a critical bottleneck in nutritional epidemiology. We developed an in-context learning approach using large language models (LLMs), in which a structured knowledge system (termed a skill) loads GI reference databases (~11,000 entries), expert decision rules, and error-correction heuristics into the model's context window (~300,000 tokens). The LLM performs GI assignments without scripted logic, functioning simultaneously as a semantic matching engine, numerical reasoning system, and expert curator. We validated this approach in two experiments. In Validation Study 1, the skill predicted the expert-curated US National GI Database (9,428 foods) using only European reference data, achieving within +/- 10 agreement of 73.7% without manual review - compared with 31.3% retention of previously published cosine-similarity approach. In Validation Study 2, the skill was augmented with US GIDB and applied to 1,157 European food descriptions classified using the EFSA FoodEx2 system, achieving ICC = 0.79 with the expert (weighted k = 0.65; triplicate ICC = 0.88). We then applied the skill prospectively to extend US dietary GI and GL surveillance to two additional NHANES cycles (2019-2023), identifying a continued decline in energy-adjusted glycemic load. Reproducibility was assessed through triplicate runs (temperature = 0, pinned model version). The skill architecture is described in sufficient detail to inform future applications of in-context learning for nutritional database construction.