The two kinds of sentence completer
Search for an “AI sentence completer” and you will mostly find web pages with a text box: paste the start of a sentence, press a button, and get several possible endings back. That is one kind. The other kind has no box at all. It works inside whatever you are already typing in, and shows the likely rest of the sentence in grey at the cursor as you go.
| Web sentence finisher | Inline sentence completer | |
|---|---|---|
| Where it works | On its own web page | In the app you are typing in |
| How you use it | Paste text, press a button, copy a result back | Keep typing; press Tab to accept |
| What it returns | Several alternative endings | One likely continuation, updated as you type |
| Best for | Brainstorming, stuck sentences, marketing lines | Everyday email, chat, notes and code comments |
| Your text goes to | The site's server | Depends on the tool; on-device ones keep it local |
Which one to use
Use a web finisher when you are stuck and want options to choose from: a headline, an opening line, a closing sentence you have rewritten five times. Use an inline completer for everything else, because the cost of a web tool is the round trip. Copying text out of an email, into a website and back again takes longer than typing the sentence, which is why sentence completion only becomes a daily habit once it lives where you write.
How to complete sentences with AI on a Mac
The quickest free option is already on your Mac: inline predictive text suggests the next word or two in grey, and Space accepts it. For whole-sentence completion that uses the conversation around the field, a system-wide app does the job. With WeType, the steps are:
Install the app and grant two permissions
Download WeType, drag it into Applications, and allow Accessibility and Input Monitoring when it asks. They let it read the field you are typing in and notice the accept key.Download a model
Pick one in Settings › Models. The default is a good start; this guide explains the others. The model runs on your Mac, so this is the only large download.Start a sentence and pause
Type in any app, in Mail, Slack, Notes, a browser or an editor, with the cursor at the end of the line. After a moment the rest of the sentence appears in grey.Take it a word at a time
Tab accepts the next word, the key above Tab accepts the whole suggestion, and Esc dismisses it. Or just keep typing.
What completions look like
A few illustrations of the kind of continuation a small on-device model produces. The suggestion is the grey part after the cursor.
#launch · Is the build ready for QA?
Almost. I just need to fix the last failing test and then I'll share it.
Things to bring on Saturday: sunscreen, water bottles, snacks and a spare phone charger.
git commit -m "Fix the date picker so it no longer resets when the month changes"
Notice what the model is doing: it is not being clever, it is being likely. That is exactly what makes it useful for the predictable parts of writing, and why it is no substitute for deciding what you mean.
Getting better completions
- Give it a head start. A completer can only continue what is there. The first clause of a sentence carries most of the information about where it is going.
- Let it see the conversation. Completions fit much better when the model can read the message you are replying to. In WeType that is the “surrounding context” setting, which needs a model of about 1.5 GB or more.
- Pin the language if you write in more than one. WeType can hold completions to any of 15 languages instead of guessing on a short field.
- Steer it with an instruction such as “keep it short” or “British spelling”. Larger models follow these more closely.
- Accept partially. Taking three words and typing your own ending is often the best use of a suggestion.
What it cannot do
An AI sentence completer predicts; it does not know. It will suggest a meeting time that sounds right but is wrong, or a name from the wrong thread. It drifts toward common phrasing, which is fine in a status update and bland in an essay. And it only works where it can read the text, which on a Mac means real text fields rather than apps that draw their own text on a canvas.