Good morning!Here's what's happening in AI today:
Meta launched Pocket, a vibe-coded gaming app
Anthropic raised Claude API rate limits 5x for all developers
CapCut rolled out Dreamina Seedance 2.0 4K across the US
Meta launched Pocket, a vibe-coded gaming app
PLUS: How to prompt Claude Sonnet 5 the right way
4 new AI tools worth trying today
AI AGENTS
Meta quietly launched Pocket, a vibe-coded gaming app where users generate and share interactive mini games from text prompts, available on iOS and Google Play with no official announcement from Meta.
Pocket is built on Gizmo, a vibe-coding app Meta acquired, and lets users describe a game in plain text which the app generates and adds to a scrollable social feed of games other users have made.
The app uses no-code AI game generation, meaning users describe the mechanics, theme, and interactions they want in natural language and Pocket builds a playable mini game from the description without any programming required.
Gizmo, the predecessor product Pocket is built on, had 635,000 lifetime installs before the acquisition with a 98% positive sentiment rating across user reviews.
Pocket is available on iOS and Google Play, though Meta has made no official public announcement about the launch.
Vibe coding has been a developer concept. Pocket turns it into a consumer behavior. Anyone who can describe a game can now make one and share it in a feed designed for discovery, testing whether AI-generated content can sustain a social loop the same way user-generated content did for early Instagram and TikTok.
AI VIDEO
CapCut rolled out Dreamina Seedance 2.0 in 4K across the United States, bringing its most advanced AI video generation model to the mobile app, web platform, and desktop across all major CapCut surfaces.
Seedance 2.0 generates 4K video from text prompts without requiring reference images, with realistic textures, movement, and lighting across a full range of visual perspectives and camera angles built into the model.
The rollout covers every major CapCut surface in the US: the mobile app across AI Lab, AI Generator, and AI Video; the web platform's Video Studio; and the desktop's AI Video and Edit Pilot tools.
Seedance 2.0 applies an invisible watermark to all generated content for traceability and blocks generation from real faces and unauthorized IP, addressing the copyright concerns that delayed its earlier global rollout.
CapCut has over one billion users globally, making this one of the broadest AI video generation rollouts to any consumer platform available.
Most AI video tools are either expensive, limited, or locked behind developer APIs. Seedance 2.0 4K lands inside an app already on the phones of a billion people, with a free tier included. For creators who want to generate or enhance video without switching tools, this removes the last reason to go elsewhere.
AI PRODUCTIVITY
Anthropic announced that it has raised Claude Platform API rate limits for all users and simplified the tier structure, which is no longer based on API spend.
The new Scale tier now allows 10,000 requests per minute, 10 million input tokens per minute, and 2 million output tokens per minute, with the latest Sonnet and Haiku models receiving 5x higher limits at the highest tier than before.
Tiers are no longer determined by how much a developer spends. Each tier now offers the same requests per minute and token throughput across the latest Opus, Sonnet, and Haiku models, so developers can choose the right model for the task without making trade-offs for rate limiting.
Advancement through tiers is automatic based on usage. Developers who need a higher rate limit immediately can request one directly from the Claude Console without going through a manual approval process.
The simplified three-tier structure (Start, Build, Scale) applies across the native Claude Platform, Claude Platform on AWS, and Claude in Microsoft Foundry on Azure.
TRate limits have been one of the most consistent friction points for developers building production applications on Claude. Decoupling tiers from spend means developers can scale their usage without worrying about whether their budget unlocks the throughput they need. For teams building agentic workflows that fire many parallel requests, the 5x increase at the Scale tier removes one of the last real infrastructure reasons to route around Claude.
HOW TO AI
I opened my saved prompts folder last week. Two years of 500-word instructions telling Claude how to think, what to check, what role to play.
Every single one written for a model that no longer exists as your default.
Claude Sonnet 5 dropped June 30. If you're on Free or Pro, you're already on it. And the way it works has changed enough that your old prompts are now slowing it down.
What changed under the hood
Three things are built into Sonnet 5 that you used to have to ask for:
Adaptive thinking is on by default. It decides how deeply to reason based on your task. You don't prompt it to think. It already is.
It checks its own output without being asked. Anthropic confirmed this in the launch notes. Zapier's team tested it: "It finished end to end. That used to stall halfway."
It plans its own steps. Tasks that Sonnet 4.6 would stop halfway through and hand back, Sonnet 5 completes.
When you write a prompt that micromanages all of this, you're fighting the model. Your saved prompts were the right call for the old model. They're overhead now.
What to delete
→ "You are an expert in..." It already knows the domain. This adds tokens, nothing else.
→ "Think step by step." Adaptive thinking is on. You're telling it to breathe while it's running.
→ "Step 1. Step 2. Step 3..." Give it the goal. It plans better than your numbered list.
→ "Re-read your response before answering." Built in. This is one of Sonnet 5's core behavior changes.
→ "Do not hallucinate." This never worked. And Sonnet 5's hallucination rate is lower by design.
The 4-line structure that replaces it
[The situation: everything you know, dumped raw. Who's involved,
what happened, what you're working with, since when.]
Goal: [the outcome you want, not the steps to get there]
Use /[your-skill] for the [deliverable].
Ask me what you don't know before you start.The situation gives Claude everything. The goal tells it where to land. The skill handles your format and preferences. The last line stops it from guessing.
That last line matters because Sonnet 5's lower sycophancy means it'll actually ask rather than proceed with assumptions.
Where to start
Open your most-used saved prompt. Find every line that tells Claude how to think rather than what to do. Delete them. Run the prompt.
Most people find the output stays the same or improves.
The model got smarter. The prompts should get shorter.
The full guide is available for paid subscribers here.
The full guide covers the before/after comparison, the exact screenshot proof, and what specifically to audit in your saved workflow folder.
Meituan launched LongCat-2.0, a 1.6 trillion parameter mixture-of-experts model with 48 billion active parameters and a 1 million token context window, now available on OpenRouter with benchmark scores competitive with leading frontier models including 59.5 on SWE-bench Pro.
Synthetic Sciences released OpenScience, an open-source AI research workbench that runs the full research loop in the browser including literature review, hypothesis, code, experiments, and write-up, with 290+ skills and connections to 30+ scientific databases including arXiv, PubChem, and UniProt.
OpenClaw shipped v2026.6.11, improving channel delivery reliability across Telegram, WhatsApp, Matrix, iMessage, and five other platforms, while adding Slack router relay mode for multi-gateway deployments and better model recovery when provider credentials or streams fail.
🎬 Runway: AI video generation and editing with Gen-4 models for text-to-video, image-to-video, and video-to-video.
🔍 Perplexity: AI-powered search that reads the web and returns cited answers instead of a list of links.
🎨 Gamma: Turn a topic, outline, or document into a polished presentation, document, or webpage.
⌨️ Windsurf: An AI-native code editor that understands your entire codebase.
✓ Full archive of premium guides with ready-to-use prompts
✓ Structured AI courses (step-by-step, start-to-finish)
✓ Every upcoming premium tutorial
THAT’S IT FOR TODAY
Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!
See you tomorrow :)
- Dr. Alvaro Cintas