Yazım Yardımcısı

Ten actions for copied text, written by Apple Intelligence, the on-device model. The text never leaves your Mac.

Action	Offered on	What it pastes
Fix spelling and grammar	Text	The same text with its mistakes corrected, and nothing else changed
Make shorter	Text	About half as long, with the facts kept
Make more formal	Text	The same message in a polite, professional tone
Make friendlier	Text	The same message in a warmer, more casual tone
Turn into bullet points	Text	One idea per line, each starting with `-`
Extract action items	Text	A checklist: `- [ ] Ada: send the draft by Friday`
Draft a reply	Text	A short reply to the message, in its tone
Explain this	Text and code	A plain-words explanation, at most about 150 words
Write a commit message	A copied diff	A Conventional Commits message
Write a PR description	A copied diff	`## Summary` and `## Changes` in Markdown

Rewrites answer in the language of the text, and never translate it: that is what Translate is for. Translation and summaries are not repeated here.

Examples

Copy can u send me the report before the meting tmrw, press ⇧⌘V, press ⇧ on the item and choose Fix spelling and grammar. It pastes:

Can you send me the report before the meeting tomorrow?

Copy the output of git diff --staged and choose Write a commit message:

fix(upload): retry a failed request up to three times

A short network drop no longer fails the whole upload: the client
retries with a growing pause between tries.

Short text at once, long text as a job

An action has 10 seconds, and rewriting a long text takes the model longer than that. So:

Fix, Make shorter, Make more formal, Make friendlier, Bullet points, Action items and Draft a reply come in two forms, and the panel offers only the one that fits the item:
- Up to 400 characters (a chat message, a short mail), the action runs at once and pastes, like any other. In Chinese, Japanese, Korean or Thai, where a character takes the model about as long as a short English word, that is fewer than 150 of those characters.
- Longer text runs as a job. The panel closes, and a card under the menu bar icon shows the progress ("2 / 5 parts") with a button to cancel. When it is done, the result is pasted if the app you started from is still in front; otherwise it is on the clipboard.
Explain this, Write a commit message and Write a PR description always run as jobs: an explanation or a description is as long for one line as for a page.

Characters are counted on the trimmed text, as the app counts them: some emoji, such as a flag or a thumbs-up with a skin tone, count as two or more.

On a Mac with an M4 Pro chip, each of the seven at-once actions answered a text at that limit in 0.3 to 2.3 seconds, in English, Turkish, Chinese and eleven other languages. A slower Mac takes longer, and the 8.5-second notice below covers the rest.

The panel shows one row for each task. Settings → Extensions lists every command, so the seven tasks with two forms appear there twice, with the same title. When you make one of them your ⇧↵ default, it is the default for that form only: for text up to 400 characters, or for longer text, depending on the item you set it on. Set it on a short and on a long item to have it for both.

Long texts

The model reads a few thousand words at most, so a long text is prepared first:

Fix, Make shorter, Make more formal, Make friendlier, Bullet points, Action items work part by part. The text is cut at paragraph breaks, then at sentence ends, then between words, and each part is rewritten on its own; the line breaks between parts are kept as they were. Action items from all parts make one checklist, without repeats. These take up to 40,000 characters (about 6,000 words); a longer text is refused with a notice, never cut.
Draft a reply reads the newest message of a copied mail thread: it stops at a line such as On Mon, 3 May 2026, Ada wrote: (also when the mail wrapped it over two lines), the Turkish Apple Mail line, -----Original Message----- or a mail client's From: / Sent: header, and leaves out lines quoted with >. A text still longer than 5,000 characters keeps its start and its end.
Explain this keeps the start and the end of a text longer than 5,000 characters.
Commit message and PR description get a condensed diff of at most 6,000 characters: a list of the files with their added and removed lines, then the changes. Lock files (package-lock.json, yarn.lock, Cargo.lock and others), minified files and binaries are counted but not shown. Context lines go first when the diff is too long, then each file keeps an even share of the room. Lines over 300 characters are cut. A diff from git diff, diff -u or svn diff works, with one file or many.

The model holds about 4,096 tokens at a time, shared by its instructions, the text and the answer. A token is about three or four characters of English, fewer of most other languages, and about one character of Chinese, Japanese or Korean. So the parts are measured in tokens as well as characters: a text in those languages is cut into shorter parts, and every answer has room to be a quarter longer than its part. If a fix, a shorter text or a new tone still comes back cut short, nothing is pasted and a notice says so, rather than pasting the text without its end. An answer counts as cut short when it is far shorter than its part, or when it is shorter and stops in the middle of a sentence where its part ends one.

Languages

Rewrites, lists, replies and explanations of text are in the text's own language. The extension names that language to the model, because the model answers in the language of its instructions otherwise: asked only to keep "the same language as the text", it rewrote German, French, Spanish, Dutch, Japanese, Korean and Chinese texts in English when asked for a formal or a friendly tone. Named, it kept every one.
Every language Apple Intelligence speaks on macOS 26 is recognized: Chinese, Japanese and Korean by their script, Vietnamese by its letters, and English, Turkish, German, Dutch, French, Spanish, Italian, Portuguese, Swedish, Danish and Norwegian by their common words and letters. When two are about as likely, as Danish and Norwegian often are, both are named, and the model keeps whichever it is. Chinese keeps its simplified or traditional characters.
A language is named only when the text clearly is in it: a German or Dutch text is not taken for English, nor an Azerbaijani one for Turkish, nor a Catalan one for Spanish, because naming the wrong language would have the model translate your text. Then the model is asked to keep the text's own language.
Explain this on code answers in the app's language, also when the code was copied as text.
Commit messages and PR descriptions are in English, the convention for both.

What is cleaned from the answer

A small model sometimes wraps its answer. Before pasting, the extension takes off a code fence around the whole answer, quotes around all of it, a first line such as "Here is the corrected text:" or "İşte düzeltilmiş metin:", and the markers the text was sent in. Each is kept when your own text had it, and a first line that rewrites your own ("Here are my thoughts:" for your "my thoughts:") stays.

Bullet points become - lines, numbered or not; action items become - [ ] lines.
A commit message's subject is at most 72 characters, with its type in lower case and no full stop, and its first verb in the imperative ("Added retry logic" becomes "add retry logic"). A feat that only fixes becomes fix, and a change to documentation or tests alone becomes docs or test. The body is wrapped at 72 characters, without lines that only repeat what the diff adds and without code blocks.
A rewrite loses a signature the model added at its end, such as "Best regards," over "[Your Name]", when your text had none. A reply keeps its placeholders: they mark what only you can fill in.
Square brackets the model escaped (\[name]) are made plain again, unless your text has backslashes of its own.
A PR description's "Summary:" and "Changes:" become ## Summary and ## Changes.
When the answer is the model declining ("I'm sorry, but I can't help with that") and has nothing to do with your text, nothing is pasted and a notice says so. A rewrite of your own "sorry, I can't make it" is a rewrite, and is pasted.
When Fix gives back your text exactly as it was, the notice says "No mistakes found." A fix of only a capital letter or a full stop is a fix, and is pasted.

When Apple Intelligence can't answer

Apple Intelligence must be turned on in System Settings, on a Mac that supports it, in a language it supports.

The action that runs at once shows the app's own message as a notice ("Apple Intelligence is turned off…"). It does not fail: a failed action counts toward turning an extension off, and this is not the extension's fault. If the model takes more than 8.5 seconds, a notice says so, before the 10-second limit.
A job fails on its card, with the app's own message.
Cancelling a job stops it at once.

Permissions

None. The on-device model is a service of the app itself: the extension needs no network, runs no programs and reads no history.

Privacy

Your text goes only to the on-device model, and only when you choose an action. It is never logged and never put in a notice. It goes into the model's prompt between markers it cannot close, never into the instructions: Apple's model obeys its instructions over the prompt, so text placed there could steer it.

Limits

The on-device model is small. It can miss a mistake, misread a tone or describe a diff loosely. Read the result before you send it.
Parts of a long text are rewritten one by one, so the tone can shift a little from one part to the next.
The seven tasks with two forms are listed twice in Settings, and a ⇧↵ default applies to one form (see above).
Languages are told apart by their common words and letters. A very short text in a language other than English or Turkish, such as a one-line Swedish question, may get no language named; the model is then asked to keep the text's own, and a change of tone can come back in English. A text that mixes two languages evenly gets both named.
A rewrite that stops early right at the end of a sentence, and is not far shorter than its part, can't be told from a finished one.
A diff is recognized by a diff --git line or a hunk header (@@ -12,7 +12,8 @@) at the start of a line, within its first 4,000 characters.
A commit subject the model writes longer than 72 characters is cut between words.
Needs yabbistir 0.4.0 or later.

Sources

Apple, TN3193: Managing the on-device foundation model's context window: the 4,096-token context shared by instructions, prompt and answer; about three to four characters a token in English and about one in Chinese, Japanese and Korean; splitting a large task into several calls; why a tight answer limit can cut an answer short.
Apple, Improving the safety of generative model output: instructions take precedence over the prompt, so untrusted text stays out of them.
Apple, Supporting languages and locales with Foundation Models: the model answers in the language of its input; the exact phrase "The person's locale is tr_TR."; "You MUST respond in …".
Conventional Commits 1.0.0: the type(scope): description form, ! for a breaking change, and the types it names.

The word lists in src/language.ts and src/words.ts, and the list of lock files in src/diff.ts, are written for this extension.

Develop

npm test -- dev.yabbistir.ai-writing          # the tests, against the fake host
npm run build -- dev.yabbistir.ai-writing     # build, check and pack dist/dev.yabbistir.ai-writing-1.0.0.yabext
npm run dev -- dev.yabbistir.ai-writing       # build, link into yabbistir and rebuild on every change

Neler yapabilir