Built with Claude: Japan 2026 (November 2–10, 2026, online), organized by Anthropic and Cerebral Valley, was Anthropic's first nationwide hackathon in Japan. From over 5,500 applicants, 250 were selected to participate (an acceptance rate of under 5%), and 225 projects were submitted. This work was designed, implemented, and submitted single-handedly during the nine-day event.
The project addresses a problem faced by working family caregivers and older adults: people cannot ask about public support programs they do not know exist. Because public programs require an application, simply not knowing about them is itself a barrier. With this system, users describe their situation in free text and answer three multiple-choice questions. The system then presents relevant programs, including ones the user did not ask about, together with the order of procedures, rationale, where to apply, required documents, and sources.
Technically, the work applies a Rules as Code approach, in which program requirements are expressed as machine-executable rules, and makes three contributions.
(1) Eligibility is not left to LLM output. It is decided by rules that Claude Code wrote and tested from official municipal web pages, which keeps the results reproducible and explainable. Manually encoding policy documents has been a major cost in Rules as Code; this work shows how an LLM agent can take on that step.
(2) Interpretation of free text is separated from evaluation of program requirements. When Claude extracts attributes such as age and living situation, it also shows the supporting phrases, so users can check how each conclusion was reached.
(3) The service covers all 1,741 municipalities in Japan, with 8,453 support programs collected and structured from prefectural and municipal websites.
The system reached nationwide operational coverage within the event period and provides a concrete example of using generative AI to build Rules as Code for improving access to public-sector information.
