Your voice, not a robot’s

“Write like me” is too vague for a model to hold on to. So Shiva writes the voice down as a style card: how warm, how formal, how long the sentences, how much Hindi in the mix, which script, emoji or not, the owner’s own sign-off and phrases, words never to use, and the business’s stance on things like refunds.

Each of these is a separate dial, which research found gives finer control than one “write like X” instruction.

In setup the owner pastes a few replies they have written before, and ticks a box agreeing they may be used as examples. Shiva keeps up to ten.

With fewer than three, there isn’t enough to measure a voice against, and the voice check says “not run” rather than guessing.

Here is a reply to Rohit as a generic AI tool might write it. It is polite, and it could have come from any hotel anywhere.

Shiva’s rule check flags the stock phrases, like “we are thrilled” and “world-class”, and the voice score sits below the line. A study of Indian and American writers found AI suggestions pulled Indian writers toward Western styles; this is the pull Shiva pushes back on.

The same answer in Meera’s own way: short sentences, thanks first, plain words, her “Do come back”, her sign-off.

No stock phrases, and the voice score is well above the line. Both replies on this page were written by hand to show the difference; the scores are computed live by Shiva’s own voice check.

The voice check ignores what a reply is about and looks at how it is written: sentence length, commas and exclamation marks, capitals, emoji, script, and how often common little words like “so”, “very”, “we” and “ji” appear. Measuring style without topic is what makes this a voice score, not a topic score.

The score is shown to the owner and never blocks Approve. Its line is a starting choice that will be tuned on real owners’ replies.

An old reply might mention an old price. Past replies are kept as voice examples only, and the draft can’t cite them. If a draft copied “₹2,800” from one, the fact check would block it, because the fact sheet now says ₹3,200.

The card is also sent again with every part of the draft, because a model slowly slips back to a generic voice within a few turns of a long chat.

Every owner has words they dislike: “property”, “valued guest”, a competitor’s name. Put them on the card and any draft that uses one is blocked until it is fixed.

Try “soon” on the stage: Meera’s own reply uses it, and the rule check catches it at once. Next: approving with a tap, and the receipt it leaves.