How the AI stories are written, and checked.

Every AI narrative starts from structured match state, never from a blank prompt. Before a single word is generated, the system assembles a precise snapshot: the score, the overs bowled, the batters at the crease, the bowling figures, the fall of wickets, the required rate, and the momentum moments the engine has already detected. The model is asked to describe that snapshot, not to invent one.

The model and the voice

Narratives are generated with Claude, instructed to write with the clarity of Harsha Bhogle and the analytical sharpness of Jarrod Kimber. The system prompt is strict about voice. It bans clichés outright: there are no thrilling encounters, no nail-biting finishes, no fake hype. It bans em dashes. It requires the writing to lead with the most decisive moment, to use specific numbers, and to end with a concrete thing to watch for next.

The guardrails

The model is given data, not opinions, and it is told to stay inside that data. It cannot reference a statistic that was not supplied. It cannot name a player who is not in the match state. When the model is uncertain, the instruction is to be specific about what is known rather than to fill the gap with confident invention. Every story is capped at roughly 200 words, because a longer narrative is usually a sign the model has started to pad.

The fact-check

Generated stories are validated against the same structured state that produced them. Scores, names and numbers are checked for consistency before a story is shown. If the live data feed and the narrative disagree, the data wins and the story is regenerated. For the daily digest, results are additionally cross-referenced against ESPNcricinfo before publication. A story that cannot be verified is not published.

Eleven languages

Narratives are generated in eleven Indian languages: English, Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi and Odia. Each is generated from the same match state, in parallel, not translated in sequence. A translated sentence inherits the rhythm of its source language, while a natively generated one reads the way a person in that language would actually write. The guardrails apply identically to all of them.

The limits

OverByOver is honest about what the AI does not do. It does not predict results. It does not assign blame. It does not editorialise beyond what the numbers support. When an AI key is not connected, the platform falls back to hand-written editorial copy rather than showing a broken or generic story. The goal is not to sound authoritative. It is to be correct, and to be useful.