Blog

Stop Retyping the Sentence You Wrote Yesterday

Here's a number I've never been able to find a reliable source for, so I'll just tell you my guess: most of what you retype all day isn't actually the same sentence you typed before. It is the same shape, wearing different clothes. And that's why the obvious fix, saving snippets of the sentences you repeat, never quite works, and why the less obvious fix, letting a model finish them, does.

I want to show you the difference, because I think it's the single most useful idea in this whole category of software, and almost nobody explains it.

The two kinds of "I typed this before"

When people notice they're retyping, the instinct is to save the sentence and reuse it. The address. The standard reply. The paragraph that has not changed in two years. That works, for a while, and then it stops, because most of what you retype isn't that kind of sentence.

The follow-up email you write every week isn't the same email. The date changes, the project changes, the ask changes. The only thing that stays the same is the first four words and the rhythm. The status update is the same shape every Friday, but the content is always new. The reply that starts "that makes sense" goes a dozen different directions after those three words. These are near-repeats, and they're the quiet majority of what a working writer types.

Why snippets fail at near-repeats

Here is the trap. You make a snippet for the follow-up email, and it's perfect for exactly one version of that email, the version you wrote when you made the snippet. Then the situation changes, as it always does, and the snippet is close but wrong, and you spend more time editing the snippet's output than you'd have spent typing. So you make another snippet for the variation, and then another, and before long you've a library that's really a museum of every email you've ever sent, and maintaining it's a second job.

A snippet is a recording. It plays back exactly what you recorded, and a near-repeat isn't a recording, it's a live performance with the same opening riff every night. I wrote about the difference in the TextExpander post, and the short version is that boilerplate wants to be saved and near-repeats want to be predicted.

Why a model can handle it

Here is what a model does differently. It doesn't save your sentence. It learns the shape of how you write, the rhythm, the openings you favor, the way you close, and then it predicts the next few words based on that shape, in the moment, for the specific sentence you're writing right now. The date is new, so the suggestion includes the new date. The project is different, so the suggestion follows the new project. It isn't replaying yesterday's email. It is finishing today's email the way you'd have, because it has learned how you finish emails.

That gray text isn't pulled from a library. It is generated for this sentence, in your voice, based on the thousands of sentences you've typed before it. The model learned that you say "I'll review" more than "let me look at", that you favor "by noon" over "later today", and it applies that knowledge to the sentence in front of you. It is the difference between a parrot and a band that knows your favorite songs.

The honest limits

I'm not going to oversell this, because the near-repeat insight has edges. A model handles the sentences where your writing is consistent, the openings, the closings, the shapes you use daily. It is less useful on the sentences where you're genuinely thinking, the ones where the content is new and the shape is too. If you write something truly novel, autocomplete has less to grab onto, and it should stay quiet rather than guess. That is the suppression principle doing its job.

And there's a second honest note. The model only learns from what it has seen, which means the first few days are weaker than the first few weeks. Your writing history builds up, the suggestions get more you-shaped, and you can read exactly what it has learned or wipe it any time. The day-one-day-three post covers the arc, and it applies here: the near-repeat magic shows up once the history has your rhythm.

What to do about it

The practical version of this is simple. Stop trying to save the sentences that vary. Let the tool that learns your voice handle them, keep the snippet expander for the sentences that truly never change, and stop treating every repeated sentence as if it were the same repeated sentence.

The trial is free for 30 days if you want to feel the near-repeat difference in your own typing. Give it a week so the history has your rhythm, watch how often the gray words finish the opening you were about to type, and notice that they're never quite the same, because your sentences never quite are. That isn't a bug. That is the whole point.

On macOS 26+ Macs, Apple Intelligence mode stays free even after the trial ends — you always keep a working path to suggestions.

Written by Amit Ashwini, who builds WriteAmp and runs its marketing. More: why the Tab key beats the chat box · mini, midi, and max compared · benchmark methodology.