변경 로그

Some of you ran into a really nasty one: after a long building session, the AI would suddenly reply with jumbled text - bits and pieces of earlier messages, file names, error banners, all stitched together into something that made no sense. Once it started, it usually kept happening on every message in that chat, and the only thing that worked was starting a new chat.

This one was hiding deep in the plumbing and it took a while to track down. It's now fixed, and new replies in every chat will be clean from here on.

We also went back and cleared out the corrupted content from older messages across the database so it can't influence new replies. If you scroll back through an old chat you might see a message or two that looks empty where the garbled text used to be - that's the old corruption being cleaned up, not something new breaking. Your projects, files, and chat history are otherwise untouched.

We're really sorry to anyone who hit this. It was confusing, it wasted your time, and it stayed broken for longer than we'd have liked. Thank you to the people who took the time to report it in chat with screenshots - that's what made it possible to find. If anything still feels off after this update, please ping us and we'll jump on it.

We tuned Smart mode to make the AI feel more like a real designer working on your project. You should notice it right away.

• Cleaner, more polished results. Pages come out looking less like a template and more like something a designer actually thought about - better spacing, hover states, mobile layouts, the small details that make a site feel real.
• Stays focused on what you asked for. The AI is better at making the change you wanted without wandering off and rewriting things you didn't ask it to touch.
• Long building sessions go further. You can keep iterating in the same chat for longer before anything starts to slow down or get confused.
• Stops repeating itself. The AI was sometimes reading the same file two or three times in a single change. We fixed that, so changes are faster and use noticeably fewer credits.

Nothing in your workflow changes. Same chat, same buttons. Just open a project and try a redesign - you should feel the difference on the first turn.

Fixed: Voice Transcription Was Stuck on Russian

When you used the microphone in chat, your voice was being transcribed as Russian no matter what language you actually spoke - so English speakers got back garbled, transliterated text. The default was wrong and was getting saved permanently in your browser, which is why even people who never touched the setting kept seeing it.

Voice transcription now auto-detects your language from the audio. Speak English, Spanish, German, Japanese, anything - it just works. The fix runs across the whole stack, so it applies to every existing user automatically the next time you open the app.

New: Transcription Language Menu

Want to lock voice input to a specific language instead of auto-detecting? There's now a dedicated Preferences → Transcription → Language menu with all 10 supported languages plus Auto-detect. Picking a specific language can improve accuracy if you have a strong accent or if you switch between languages a lot and want voice input to stay consistent.

Heads up: as part of this fix, your chat preferences (selected AI model, custom instructions, open chat tabs) will reset to defaults the first time you open the app after this update. Your projects, files, and chat history are not affected. If you notice anything off, just refresh the page - and if something still feels wrong, ping us in chat and we'll sort it out fast.

Fixed: Multi-File Projects Now Build Correctly

When AI built websites with multiple files (HTML, CSS, JavaScript), it could sometimes produce files that didn't match each other - for example, CSS class names that didn't correspond to the HTML. This happened because the AI lost visibility of files it had just written. Now the AI always sees the full content of its recent work, so multi-file projects stay consistent.

Fixed: Conversations No Longer Get Stuck

In rare cases, if the AI took a full-page screenshot of a very tall website, the conversation could become permanently stuck - every message would fail with "Something unexpected happened." This has been fixed. Screenshots are now automatically resized to safe dimensions, and if an issue does occur, you'll get a clear message explaining what happened.

Improved: Web Research for Large Pages

The AI can now fetch content from larger websites (up to 5 MB). Pages that previously failed due to size limits - like sites with inline JavaScript - are now handled automatically. If a page is too large for direct fetching, the AI falls back to an intelligent content extraction service that pulls just the useful text.

Improved: AI Research Tasks

When the AI researches websites to gather content for your project, it now has more room to work - up to 20 steps instead of 8. This means it can thoroughly explore multi-page websites without running out of steps halfway through.

Projects with long chat histories and lots of files now load significantly faster. Instead of downloading every message on login, chats now load on demand - only when you open them.

• Instant project loading - Login and project switching no longer wait for chat history to download. You get to your code immediately.
• Smart message loading - Chat messages load when you open a conversation, with older messages fetched as you scroll up.
• Lower memory usage - Only active chats consume memory, so even projects with hundreds of AI conversations stay snappy.

Fixed: AI Repetition Loops

In rare cases, AI models (especially Gemini Flash) could get stuck in a loop, endlessly repeating the same text. This wasted credits and could freeze your browser with a gigantic message. We've shipped a multi-layer fix:

• Automatic loop detection - The AI now detects when it starts repeating itself and stops immediately, saving your credits.
• Tuned model settings - Adjusted temperature and repetition penalty parameters to prevent loops from starting in the first place.
• Large message protection - If a response is unusually long, it's now truncated in the UI with a "Show full message" button, preventing browser freezes.

Fixed: Image Uploads in AI Chat

Over the past few days, some of you experienced issues when uploading images to AI chat - conversations would freeze or show an error message. We've been investigating this intensively and shipped multiple fixes:

• CDN compatibility fix - Google's AI servers couldn't fetch images through our CDN. We now route image URLs through a direct path that bypasses the issue entirely.
• Long conversation support - Conversations with many images (10+) would fail because of a provider limit on image URLs per request. We now intelligently manage the image budget across your conversation history, keeping the most recent images while preserving text references to older ones.

If you were affected, your existing chats with images should work again. Thank you for your patience while we tracked this down.

Improved: AI Clarifying Questions

The AI's question wizard now supports two types of questions: multiple choice (A/B/C/D) for decisions, and text input for open-ended questions like "What's your business name?" Previously, the AI was forced to turn everything into multiple choice, which sometimes caused errors or awkward options. Now it picks the right format automatically.

Playcode is now available in 10 languages - both the marketing site and the editor itself. Every menu, tooltip, button, and AI chat interface has been translated.

Your language is auto-detected from your browser. You can also change it anytime in Menu > Preferences > Language.

The same announcement in all 10 languages:

• EN: Playcode now speaks your language! Editor, menus, and AI chat - all translated.
• ES: Playcode ahora habla tu idioma! Editor, menus y chat AI - todo traducido.
• PT: Playcode agora fala seu idioma! Editor, menus e chat AI - tudo traduzido.
• DE: Playcode spricht jetzt deine Sprache! Editor, Menus und AI-Chat - alles ubersetzt.
• FR: Playcode parle maintenant ta langue ! Editeur, menus et chat AI - tout traduit.
• TR: Playcode artik senin dilini konusuyor! Editor, menuler ve AI sohbet - hepsi cevrildi.
• RU: Playcode теперь говорит на твоем языке! Редактор, меню и AI-чат - все переведено.
• JA: Playcodeがあなたの言語に対応! エディター、メニュー、AIチャット - 全て翻訳済み。
• KO: Playcode가 이제 한국어를 지원합니다! 에디터, 메뉴, AI 채팅 - 모두 번역 완료.
• IT: Playcode ora parla la tua lingua! Editor, menu e chat AI - tutto tradotto.

New

• GPT-5.4 - OpenAI's latest model is now available as a quality model option.
• 3 AI Modes - Economy (Gemini only, lowest cost), Smart (Gemini for reading + Claude for building), and Pro (Claude for everything).

Improved

• Help Center rewritten - All help articles updated to reflect current AI features, modes, and model names.

New

• "Undo from here" button - Every message you send to AI now shows an "Undo from here" button below it. Click it to undo all the AI's changes after that point and try again with different instructions. No more endlessly asking AI to fix its own mistakes.
• "Undo & resend" confirmation - When you undo, a clear confirmation dialog explains exactly what will happen before any changes are made. Edit your message if you want, then send it again.
• Checkpoints help page - New help article explaining how checkpoints work, when to use them, and how they relate to publishing.

Improved

• Versions renamed to Checkpoints - "Versions" was confusing for non-technical users. They're now called "Checkpoints" everywhere - menu, labels, help pages. Save a checkpoint before big changes, come back to it anytime.
• Checkpoints moved to top-level menu - Checkpoints were buried three levels deep (Menu > Project > Versions). Now they're right in the main menu, one click away. Includes a "How it works" link to the help page.
• My Projects redesigned - The "My Projects" menu item now has a distinct visual style with a project count badge, so you always know where to find your other projects.
• AI suggests undo and checkpoints - When things go wrong, the AI now suggests using "Undo from here" instead of trying to fix its own mistakes. It also recommends saving checkpoints before major changes.

Fixed

• CSS url() on published sites - Images and fonts referenced with url() in CSS (like background images) now work correctly on published sites. Previously, these paths broke after publishing.