Turning a labour law into tested TypeScript: what an AI pair taught me about legal code

This story details the process of creating AI-powered gratuity calculators for UAE workers while ensuring strict adherence to the legal code. The author, who is not a lawyer, used Claude Code (Anthropic's coding agent) to help draft the calculators. The key lessons learned are:

1. The UAE end of service gratuity calculation is more complex than commonly perceived, with multiple articles in the Federal Decree-Law No. 33 of 2021 governing the rules. No rule should be included in the code unless it is directly sourced from the official text.

2. Separating the law from conventions helps differentiate between legal requirements and commonly used conventions. The code includes a separate list of conventions, such as the daily wage calculation, to keep them distinct from legal rules.

3. Every test case should be worked out by hand to ensure accuracy. The author emphasizes that AI should not be allowed to generate test values through the function itself, as this would lead to incorrect results. The tests includes arithmetic calculations to verify the correctness of the implementation.

4. Date calculations are particularly prone to errors when using simple division by 365. The author implemented a function to calculate the exact number of years between two dates, accounting for leap years and ensuring the correct result in all cases.

5. Finally, the author stresses the importance of documenting what is not modeled in the calculator. Part-time workers and potential deductions by employers are not included in the current version, but this is clearly stated in the module header and on the page. Being upfront about limitations is better than silently ignoring certain cases.