Atomic questions beat crammed principles (Takuma Yoshida)
External note.com write-up by Takuma Yoshida (2026-10-29): on Kaggle Game Arena Bargaining, splitting negotiation principles into many small hosted TypeSafe Jev questions (then optionally moving calculable parts to Python) beat stuffing seven principles into one prompt.
What this guide points to
Takuma Yoshida published a note.com article on 2026-10-29 about splitting Jev instructions into many small questions (Japanese original title kept on the source page). The series tests whether hosted Jev can be better, faster, or cheaper than a general LLM for work-like decisions by playing games. After Othello, this post switches to the Bargaining game (hidden valuations, up to ten proposals).
Earlier, appending seven negotiation principles into a single Jev question (Strategy) made play worse than a bare Basic prompt. This post instead:
- Atomic pass — ask many narrow questions per candidate deal / item / turn (up to 189 questions when 37 moves are legal), batched in one Jev call (Jev Atomic).
- Compose — feed the answer table plus the same seven principles into a second Jev Choice of the move (composition stays model-side here; TypeSafe docs often keep the last combine step in code).
- Constantize — replace questions whose answers are pure arithmetic with Python (Python Atomic), cutting cost and latency further.
Reported Bargaining results in the article (author’s API list-price estimates, not invoices): Jev Atomic vs Basic 297–40–63; vs prior Strategy 330–14–56; Python Atomic stronger still, with median decision time and per-game cost lower than Jev Atomic. The author notes atomicization helped more on Bargaining than on Othello. Next planned matchup: Claude Haiku 5.5.
This page is an unofficial handbook pointer — verify numbers and methods against the note and TypeSafe docs.
Takeaways
- Cramming principles into one question can scatter probabilities; splitting into atomic questions can reverse the outcome.
- Batch many independent Jev questions in one request when the API allows.
- Keep deterministic math in code; ask Jev only for judgment that stays fuzzy but closed-set.