Rookie grounds language generation in engine analysis rather than letting a model narrate from the board alone. For every position I extract 27 engine-derived features (evaluation swing, win-percentage delta, checks, captures, hanging pieces, mobility change, trades, positional gains), so each explanation traces back to a measurable signal.
Deterministic templates over those features produced 3.9M+ synthetic commentaries from 197K+ FEN positions, which I used to fine-tune a T5-small seq2seq model. Benchmarked against GPT-4o-mini on 1,000 held-out positions, it won 69.6% preference from a Gemini-2.5-Flash LLM judge and 73.7% from human chess players, scored on faithfulness, relevance, informativeness, clarity, and conciseness.