Informal Seminar: Neil Gordon, From Spot Forecasts to Agent Skills: AI for preparing weather forecasts and reports

Alongside the revolution in data-driven NWP models such as AIFS, and in AI systems that write code, there has been rapid progress in the capability of AI to help with meteorologists’ work of interpreting NWP output and putting it into words. Since 2023 I have followed this progress and carried out my own experiments. These began with spot forecasts from a chatbot, then moved on to ensembles and uncertainty, area forecasts from representative locations, and impact-based forecasts. That work is now shared as the open-source Impact-Based Forecast Toolkit (IBF). Three developments have been particularly important in 2026: multimodal AI systems can now interpret forecast charts, agents guided by skills can plan and carry out investigations, and small language models running locally can write routine spot forecasts. I’ll show examples of the shift from code supplying AI with predefined inputs to AI agents deciding what evidence is needed and using tools and code to gather and analyse it. I’ll also consider how these approaches might help analysts prepare ECMWF’s Forecast Daily Report.