Figma's agent lands on the canvas; Instagram builds filters from prompts; Apple sells AI headroom in $50 packs

🔥The one thing to know

🛠Tools and releases

• Effect Builder — Instagram's new chatbot turns a text prompt and optional style reference images into a custom filter for Stories, which you preview on your own photos and refine in follow-up messages before submitting. A shipped, mainstream prompt-to-asset pipeline with a review gate — a pattern worth studying if you are putting generative features into a consumer product. Live for US users over 18; The Verge's test found Instagram's prompt-rewriting option padded a one-liner into paragraphs of style detail without visibly changing the generated result. theverge.com

• Spacefast — Automattic, WordPress's owner, launched a home for vibe-coded apps: the databases, logins, file uploads, email and analytics a prototype needs to become a real product. Bring your own agent — ChatGPT, Claude, Cursor — or just connect a Git repo. It attacks the gap designers hit when a code prototype has to survive contact with real users, and it competes directly with Lovable and Emergent. Built by Batuhan İçöz during a company hackathon and soft-launched in September, with İçöz arguing its edge is letting users pick their own agent. techcrunch.com

• Intelligence Usage Packs — Apple now lets you buy more AI usage in one-time $50 packs once you hit the monthly cap, starting with its Creator Studio apps. The first live answer to how AI features in creative tools get priced when free limits bite — expect caps and top-ups to reach design tools too. Apple had already said it would eventually charge for expanded use of select AI features; the packs are that promise turning into a product. techmeme.com

• Lovable — Lovable has hired the team behind Radish AI, a platform for running and scaling agents, to build the agent features businesses need as Lovable-built apps start running real operations. A clear signal of direction: from making apps to making apps that run a company — which changes the kinds of products designers will be asked to build on it. The three engineers split between Stockholm and a newly opened London office, with former Meta lead Nirman Gupta heading engineering for business agents. lovable.dev

🎨Designing for agents

• Same AI, opposite outcomes — the interface decides. A research round-up arguing that identical AI helps or harms learning depending on the interface around it: a purpose-built tutor doubled gains in a Harvard trial, while Wharton students with a plain chatbot tutor did 17% worse on the later unaided exam than peers with no AI at all. The usable patterns: make learning an explicit mode, ask before telling, offer a hint ladder instead of an answer button, and measure unaided performance rather than engagement — engagement rose in the Wharton study while real learning fell. All three big labs already ship that mode switch: Learning mode in Claude, Study Mode in ChatGPT, Guided Learning in Gemini. gokhanmeric.com

  “AI tutors should not do the thinking for students; they should help students build it.”

• Security friction is a design argument now. A Hacker News thread claims ChatGPT Desktop forwards your SSH login credentials to remote machines, letting an agent elsewhere sign in as you — but the pull request offered as proof was closed and never merged, so the claim is unproven. The design argument underneath is the keeper: 'always ask' and 'always allow' are both what you ship when you have no risk signals; one commenter proposes cumulative risk scoring, like a card issuer that steps in only when a purchase looks off. The thread's practical mitigations double as a permission-UI checklist: hardware keys that demand a physical touch, scoped sub-accounts, and a warning about confirmation fatigue — ask too often and people stop reading. news.ycombinator.com

• recalc — An open-source Mac app, in early beta, that makes the spreadsheet a shared workspace with your coding agents: a cell starting with '=' calculates, one starting with '@' runs Claude Code or Codex in place, and the agent's plan and answers land in the grid where you can edit them. The pattern is the point — agent work arrives somewhere you can inspect and steer, a test cell placed beside an agent's 'done' turns red when it is wrong, and everything that needs you gathers in a single list. Boards are just folders of files, and commands or agents only run when you tell them to; the README warns of bugs and breaking changes. github.com

• Google Gemini — A GitHub repository documents that Google now says, in an October blog edit and updated help pages, that connected Drive files and Gmail can be routed to human reviewers — a quiet reversal of its 2023 promise that personal content would never be seen that way. The author's tests claim files are pulled in during casual conversation, without you asking for them, sometimes as full unredacted text. For anyone designing a 'connect your apps' feature, this is the cost of ambient retrieval: the consent moment has to be explicit and visible, not a help-center edit. The author stresses the behaviour is probabilistic: one clean test proves nothing. GitHub (thisisme6598/gemini_files)

💬What designers are saying

🌍The big AI stories today

• Claude Fable 5.2 — Anthropic's Fable 5.2 ships: 30% faster, 20% cheaper at $8/$40 per million tokens — the engine under the Claude tools you use. anthropic.com

• Muse Ember — Meta's Muse Ember ($0.10 per million input tokens) handles the small steps agents repeat all day — cheap enough to run in every product, Meta says. research.meta.ai

• Nvidia — Jensen Huang admitted infidelity to staff; Trump says he keeps his job — the board's probe of the alleged spying continues at the AI chip giant. businessinsider.com

• Anthropic IPO — Reuters saw emails where Anthropic sent bankers plans to raise up to ~$69B, then called the numbers inaccurate, days before its expected listing. techmeme.com

💡One takeaway