Clef-driven ANGEL HALOS AI werewolf game cuts AI costs 84%

The developer started from a scripted-clash idea: rather than asking one LLM to both talk and vote, the AI is split like a theater troupe into a director and an actor. Clef is the director — it only picks among options and returns probabilities, with three question types (yes/no, multiple choice, and a scored grade from 'innocent' to 'almost certainly dark') and a consistency that returns the same value for the same input across five calls. Gemma is the actor, delivering lines in character. The split came after an early attempt let an angel say 'I think Aria is innocent' and then vote for Aria anyway.

To keep the AI from seeing everything, each angel only remembers what entered its view, checked every 0.5 seconds and logged as individual lines such as having watched a specific player in a specific spot about 30 seconds earlier. A fallen angel's own attack is tagged secret, and the actor is told never to mention it. Word choice is also constrained: 'kill', 'corpse' and 'blood' are banned in favor of angel-world language like 'attacked' or 'returned to the light'.

The cost work followed a familiar indie worry about pay-as-you-go bills. Initially Clef was called three times per meeting and every surviving angel spoke in the second round, which would have burned the free tier quickly. The fixes — moving to clef-flash, dropping to two Clef calls, and having only the accused and accusers speak in round two — brought about an 84% reduction. A fallback browser AI keeps a meeting from freezing if the API fails. On localization, the developer chose not to translate the Clef questions, keeping them in English for all 17 languages and passing only names, memory summaries, and logs in the local language, so a question fix needs only one edit.