Grounded in macOS 27 Golden Gate, released 14 September 2026: what separates Apple Intelligence from Siri AI, how to make English-native development the system default with Writing Tools, Visual Intelligence and the ChatGPT extension, and how to fold thirty minutes of daily Japanese study into the same workflow — including where client work should stop short of third-party models.
As of: 16 September 2026 macOS 27 Golden Gate shipped on 14 September, two days before this was written. Every claim below about feature availability, language support, and regional limits is accurate as of that date. Apple changes these things in point releases. Check Apple's own documentation before making decisions that matter.
Who this is for: the second half assumes a non-Japanese engineer in Tokyo learning Japanese. The structure is symmetric, though — a Japanese speaker learning English can use it unchanged.
One release, two things everyone is conflating
Since macOS 27 Golden Gate landed on 14 September, a wave of "Apple finally got AI right" write-ups has appeared. Nearly all of them make the same mistake: they treat "Apple Intelligence" and "Siri AI" as the same thing.
For someone in Tokyo who writes English code and reads Japanese specifications, that distinction is not academic. It is exactly the line between what you can use today and what you wait until next month for.
So before the workflow, let's settle the accounting.
Four names, four different realities
| Name | What it is | Available now? | Languages & regions |
|---|---|---|---|
| Apple Intelligence | The platform capability set: Writing Tools, system translation, image cleanup, notification summaries | Yes, expanding since 2024 | 16 languages, including Japanese and Simplified Chinese |
| Siri AI | The new Siri in macOS 27, with personal context, onscreen awareness, and system-wide app actions | Yes — but English beta only | French, Japanese, Korean, Portuguese, Spanish next month. Not available in the EU or mainland China at launch |
| Visual Intelligence | Select a window on screen with ⌘⇧Space and ask Siri about it, or have it act | Yes | Follows Siri AI's language rollout |
| Model Delegation | The mechanism that lets third-party models act as Siri extensions. ChatGPT is the built-in one | ChatGPT extension available | Depends on the extension |
The conclusion is blunt: the half that actually works today is "ask in English, get English back."
That is a fact about availability, not a style preference. For developers who already write code and documentation in English, it happens to be a tailwind. Japanese support arrives in October — and in the meantime, asking about Japanese material in English tends to give steadier results than asking in Japanese directly.
Two traps worth naming
- If your Apple ID region is mainland China, or you are in the EU, Siri AI is not available to you at launch. Anything in Parts 1 and 2 that depends on it needs adjusting.
- The hardware bar is widely overstated. Apple Intelligence's baseline features run on M1. Only the advanced features require M3 or later with at least 12GB of unified memory. An M4 Max with 64GB is not an entry ticket. What it actually buys you is covered in Part 3.
Part 1: Make English-native development the system default
1. Writing Tools and Write with Siri
PR descriptions, commit messages, API docs. Each one is a context switch from "writing code" to "writing English," and then back. Writing Tools compresses that round trip into a right-click.
macOS 27 adds Write with Siri on top: rather than only rewriting what you have, you describe what you need and it drafts it. For the very common case of "I know what this function does, I just don't want to assemble the English," that is a qualitative difference.
| Case | Raw developer input | Cleaned up |
|---|---|---|
| Commit message | added caching for the user api so it loads faster and doesn't hit db too much | feat(api): add caching layer to the user endpoint to cut response time and reduce database load |
| Jira update | fixed the bug where app crashes when user is null on login screen | Fixed a crash on the login screen caused by a null user object during authentication initialization. |
| API docstring | this function takes a string and returns a date or throws error | Parses the given string into a Date. Throws a TypeError if the input is not a valid date format. |
One practical warning: do not let the model make your commit message grander than the facts. The first row above deliberately avoids phrasing like "optimize load times and reduce database I/O overhead," because that is a promise, not a description. Rewriting tools pull consistently toward the formal and the ornate, and in technical writing that is a liability. Read the output yourself and delete the adjectives.
2. Visual Intelligence: leave the error on screen
The old debugging loop was: see the error, select it, copy it, switch to a browser, paste it, ask. macOS 27 collapses that to one step.
The flow:
- Leave the stack trace in your IDE. Do not copy it.
- Press ⌘⇧Space for Visual Intelligence and select that window.
- Ask in English. Where the request warrants it, Siri asks whether to hand the context to the ChatGPT extension.
(For orientation: on the Mac, Siri AI itself lives in Spotlight — ⌘Space — and also appears in right-click menus and a dedicated Siri app, with conversations synced over iCloud. Visual Intelligence is the entry point specifically for what is on screen.)
Prompts that earn their keep:
# Case 1: locating the real cause
Look at the error trace in the selected window. Explain the root cause,
then give me a minimal fix — not a rewrite. Point out which line is
actually wrong and why the stack is misleading.# Case 2: refactor to a known pattern
Read the component in the selected window. Refactor it to MVVM with
Swift Concurrency (async/await). Keep the public API unchanged and
add inline docstrings in English.# Case 3: filling test gaps
Based on the module in the selected window, write pytest cases.
Cover the edge cases the current tests miss. List what you decided
NOT to test and why.That last clause — "List what you decided NOT to test and why" — deserves its own note. Making the model account for its own omissions is worth more than three extra test cases. The dangerous failure mode of generative tooling is not writing something wrong; it is silently leaving something out.
3. When a request leaves your Mac — and when it shouldn't
macOS 27 splits processing into three layers, and the boundaries are unusually clear:
| Layer | Where it runs | Typical use |
|---|---|---|
| On-device | Your Mac | Writing Tools rewrites, basic understanding of screen content, the Spotlight index |
| Private Cloud Compute | Apple silicon servers | Inference the on-device model can't handle. Apple states data is neither stored nor used for training |
| Third-party extensions (Model Delegation) | ChatGPT and similar | Large-scale design work, hard algorithmic reasoning. Consent is requested explicitly, every time |
For teams doing client work, the third layer is the one that belongs in written internal policy. Here is what SolanaLink actually does:
- Client code, client data, and unpublished architecture diagrams do not go to layer three. Not because OpenAI is untrustworthy, but because most confidentiality clauses in client contracts do not distinguish between recipients.
- When an external model's capability is genuinely needed, sanitize first: replace business nouns with placeholders and keep only the structure. Most architectural questions survive that intact.
- On Apple's published description, the first two layers cover the large majority of day-to-day work. Layer three is needed less often than people assume.
This article is not legal advice. For specific questions about Japan's Act on the Protection of Personal Information or your own client contracts, consult counsel.
Part 2: Thirty minutes a day, with Tokyo as the corpus
Japanese proficiency doesn't decide whether you can write code. It decides whether you get to be in the room when decisions are made. Specifications, estimates, internal approval documents — the Japanese used there is not what language schools teach.
The good news: if Chinese is your first language, you have a head start on kanji. If your English is strong, you already own a full vocabulary of grammatical terminology. Connect both to AI and you do not have to restart from the kana chart.
1. How to split the thirty minutes: 10 / 10 / 10
| Slot | Content | Tools |
|---|---|---|
| 10 min, morning | Review yesterday's captured vocabulary and kanji | Notes + Writing Tools |
| 10 min, commute | Conversation role-play / listening | AirPods + Siri |
| 10 min, evening | One grammar point, English↔Japanese contrast | Siri AI (asked in English) |
What matters is not the duration but that every piece of material is something you actually ran into today. The textbook's "Tanaka-san is a student" will never appear in your work. The 差し戻し in this morning's spec will.
2. Kanji and vocabulary: from screenshot to knowledge base
When you hit Japanese you can't read — a train ad, a supermarket label, a spec-change notice in Slack — screenshot it, then:
1Analyze the Japanese text in the selected window.
21. Give the furigana for every kanji compound.
32. Translate into English and Chinese.
43. Flag which readings are on'yomi and which are kun'yomi.
54. Give one example sentence in a business context.
65. If the Chinese and Japanese meanings of the same kanji differ, say so explicitly.Item 5 is there specifically for Chinese speakers, and it is the most important line in the prompt. The kanji advantage cuts both ways: 手紙 is letter paper in Chinese but a letter in Japanese; 勉強 means doing something reluctantly in Chinese and studying in Japanese; 大丈夫 means a fine upstanding man in Chinese and "it's fine" in Japanese. These false friends are the single most reliable source of errors for Chinese speakers — and precisely where AI helps most, provided you remember to ask.
Send the result to Notes in one tap. After a week, select the whole set and have Writing Tools organize it into a table grouped by IT terminology, business pleasantries, and ward-office bureaucratic vocabulary. You will have built a personal dictionary without noticing.
3. Grammar: use English terminology as the bridge
Chinese speakers get stuck on Japanese grammar largely because Chinese has almost no inflection — particles, conjugation, and the honorific system have no counterpart. English has partial counterparts. Asking in English grammatical terminology hands the model a coordinate system it knows extremely well.
| English concept | Japanese concept | Example prompt |
|---|---|---|
| Transitive / Intransitive | 他動詞 / 自動詞 | "Explain 開ける vs 開く. Compare to English 'raise' vs 'rise'. Give 3 examples with furigana." |
| Relative clauses | 連体修飾 | "How does Japanese modify nouns with verbs, without relative pronouns like 'which' or 'that'?" |
| Passive / Causative | 受身形 / 使役形 | "Break down 使役受身. Compare it to the English 'being made to do something'. When is it passive-aggressive in a workplace?" |
| Register / Formality | 敬語 / 丁寧語 | "Rewrite this casual email in keigo for a senior client. Label each 尊敬語 and 謙譲語 verb you used." |
"Label each … you used" in the last row is a general technique: make the model annotate its own moves and you walk away with the rule, not just the result.
4. Speaking and email: two levers already in place
Role-play on the commute. Put in AirPods, and on the Yamanote line say:
"Act as my Japanese project manager. Let's do a 5-minute standup where I report progress in Japanese. After each turn, correct my grammar and suggest a more natural phrasing."
Live Translation on AirPods now supports Japanese (iOS 26.1 or later recommended, with a supported AirPods model). Treat it as a safety net, not a study tool — it keeps a conversation from stalling when you miss something live. The learning happens afterwards, when you review what you missed.
Turning email into Japanese office register. Draft in whichever language is fastest for you, then have the system convert it:
Draft (English): I finished the backend API. Please check it. If there are bugs, let me know and I will fix it by tomorrow.
Converted (Japanese): バックエンド API の実装が完了いたしましたので、ご確認のほどよろしくお願いいたします。万が一、不具合や修正点などがございましたら、ご遠慮なくお申し付けください。明日中に対応させていただきます。
5. One limitation that has to be stated
AI-generated Japanese is not reliably correct. Its error rate on doubled honorifics (二重敬語) and on mixing spoken and written registers is clearly higher than its error rate in English.
The workable posture: use AI as a draft generator and an explainer, never as the judge. Anything in Japanese going to a client gets thirty seconds of a native colleague's attention before it is sent. That alone catches the most embarrassing category of mistake. Real progress comes from the loop — AI drafts, a human corrects, you understand why — not from any single step of it.
Part 3: Make it a system, not a matter of willpower
1. Describe a Shortcut
Shortcuts in macOS 27 gained Describe a Shortcut: state what you want in plain language and the required steps are assembled for you. What used to mean dragging twenty blocks around a GUI is now a sentence.
A morning routine you can hand it verbatim:
"Every weekday at 8am, fetch today's top posts from Hacker News and Qiita. Summarize each in English as three bullet points. From the Japanese ones, extract five N1/N2 vocabulary items with furigana and one IT-context example sentence each. Append everything to the 'Daily Standup & Study' note."
The point is not the automation. It is that the startup cost of studying drops to zero. Thirty minutes of Japanese a day fails not because thirty minutes is long, but because deciding what to study today consumes willpower on its own. If the material is waiting before you wake up, all that's left is reading it.
2. Hardware: what an M4 Max is actually buying
Back to that overstated requirement. Baseline Apple Intelligence runs on M1; advanced features want M3 or later with 12GB or more. So an M4 Max with 64GB is not unlocking features.
What it buys is concurrency. With 64GB of unified memory, a local build, a Docker environment, and several model contexts can all stay resident, and on-device inference doesn't contend with your compiler. For someone invoking Visual Intelligence dozens of times a day, the difference shows up not as can I use this but as does it make me wait.
If you are specifying a machine: put the budget into memory before the top-end chip. A 24GB M4 serves this workflow better than a 16GB M4 Max.
Closing: collapsing two habits into one
The value here isn't in any individual feature. It is that writing code and learning Japanese start using the same muscle memory.
On the development side, the friction of copying between applications disappears. An English-native output path — rewrite in Writing Tools, ask via Visual Intelligence, delegate to ChatGPT only when warranted — makes the distance between "I've worked it out" and "it's written down" as short as it goes.
On the language side, you drop the vocabulary book and convert the Japanese you actually meet each day into study material. With kanji familiarity and English grammatical terminology as bridges, AI adds to the knowledge structure you already have instead of demanding you rebuild it.
Worth staying clear-eyed about: this is early. Siri AI's Japanese arrives in October, mainland-China Apple IDs are excluded for now, and AI-written keigo still needs a human check. This workflow is a starting point, not a destination.
For someone writing code in Tokyo, though, it is a good place to start.
Note on the shelf life of this article
Feature availability, language support, and regional limits described here are accurate as of 16 September 2026, sourced from Apple's own announcements and major press coverage. Apple Intelligence features change frequently in point releases, and this article makes no guarantee of long-term accuracy. It is not legal advice; consult a qualified professional on questions about Japan's personal information protection law or your own contracts.
References
- Apple Newsroom, "Major updates for Apple's software platforms are now available" (September 2026): https://www.apple.com/newsroom/2026/09/major-updates-for-apples-software-platforms-are-now-available/
- Apple Newsroom, "Siri AI, a profoundly more capable and personal assistant, is here" (September 2026): https://www.apple.com/newsroom/2026/09/siri-ai-a-profoundly-more-capable-and-personal-assistant-is-here/
- Apple Support, "What's new in the updates for macOS 27 Golden Gate": https://support.apple.com/en-us/127257
- Apple Support, "Use Live Translation with your AirPods": https://support.apple.com/en-us/123185
- MacRumors, "macOS Golden Gate Now Available With Siri AI, Visual Intelligence and More" (14 September 2026): https://www.macrumors.com/2026/09/14/apple-releases-macos-golden-gate/
Written by Tony, engineer at SolanaLink. Comments and questions welcome.
Tags: #AppleIntelligence #DeveloperTools #JapaneseLanguage #Productivity #Tokyo
コメント
コメント (0)