Finding the right voice by branching the draft

The paragraph is fine. Grammatical, clear, on message, and it does not sound like you at all.

It's the opening of tomorrow morning's newsletter, and you've been circling it all evening. In a normal chat, the circling is the problem. You say "warmer" and the model rewrites everything, including the two phrases that were working. You say "less corporate" and it swings into fake-chummy. Each rewrite inherits the damage from the last one, the good version from a few turns back has scrolled into the fog, and after enough passes the paragraph belongs to nobody.

The trouble is that tone isn't a destination you steer toward one correction at a time. It's a choice between real alternatives, and you can't choose between things you can't see at the same time.

So stop steering. Highlight the paragraph itself, right there in the reply. A tooltip appears. Click Branch. This is the precise version of branching (there are three ways to branch; this one starts from exactly the text you selected), and the new branch opens in a pane beside the original. Ask it for the casual version. Back in the parent, highlight the same paragraph, branch again: formal this time. Once more: witty.

Now three panes sit next to each other holding three committed versions of the same paragraph, each written by a model that wasn't hedging against your last correction. The casual one is relaxed all the way down. The witty one actually risks the joke. None of them is a compromise, which is what the steering method produces every time.

Inside each branch you can keep pushing that one voice without touching the others: make the casual one shorter, ask the witty one to lose the pun but keep the energy. Corrections land on the version they were meant for. And if three voices aren't enough, branch again; you can branch as many times as you like, though in practice more than four stops being a choice and starts being a slush pile.

Then do the thing that sounds too simple to be the trick: read them aloud.

The casual one buries the announcement in throat-clearing. The formal one is a press release wearing your byline. The witty one is wrong for this audience, except for one sentence, and that sentence is the best writing on the screen.

This is also the honest limit of the exercise. GPTree can produce the three voices and hold them still while you look, but it can't tell you which one is yours. No model knows that you'd never open with "folks," or that this particular joke reads as smug to the people you write for. The branches turn taste from a vague dissatisfaction into a decision between visible options. The taste is still yours.

Harvesting the winner is mechanical. Highlight the one great sentence in the witty branch and choose Merge from the tooltip, aimed back at the parent. The AI-Assisted strategy works the line into the plain version's paragraph instead of stapling it on; you check the change in the preview, removals in red and additions in green, and confirm. If you want to weigh two versions more carefully first, the side-by-side compare renders two branches as a diff, which is surprisingly clarifying for tone: you see the exact words each voice swapped, not just the vibe. (The side-by-side compare is a Pro and Team feature. Branching and merging are on every plan; the Free plan includes 3 merges a month, and Pro and Team lift that to unlimited. Honestly, three panes plus reading aloud gets you most of the way.)

The send goes out on time: the plain version's structure, the witty branch's one earned line, an opening that finally sounds like a person. The other two branches stay in the tree. Next month, when you've forgotten what your formal register even looks like, it'll still be there, holding its version, still committed to the bit.