Figma Weave puts one prompt in front of several models at once

Weave runs one prompt past several models, then forks

Figma published a short rundown of Weave today. The two features that change a designer's routine: model comparison sends one prompt through several AI models side by side, and branches fork a workflow from the same starting point while the original stays put.

Around those sit editing nodes named Color correction, Compositor, Levels, Trim, Crop and Painter, plus a Prompt Enhancer node and a Run Any LLM node. Teammates can also run Weave tools from Figma Design. The post is a recap, so it gives no ship dates, plan tiers or credit costs per feature.

Picking a model stops being the first decision and becomes the last one.

We covered Weave tools running from chat clients in August. Comparison and branches are the new part, and they turn model choice into something you can look at.

Compare

One prompt, several models, side by side

Branch

Fork from the same start; the original is kept

Edit nodes

Color correction, Compositor, Levels, Trim, Crop, Painter

Not stated

Pricing, credits, per-feature dates

Webflow's CLI can wrap exported components in @scope

Webflow's DevLink export pulls a site's components and global styles into a React or Astro app, and its walkthrough shows a config switch that matters for anyone mixing builder CSS with their own. Turn on cssScopes and every exported rule sits inside an @scope block keyed to a generated class, which you put on the body.

webflow devlink export

// webflow.json, inside "devlink-export"
"cssScopes": true

Re-exporting overwrites local edits, so the post treats Webflow as the source of truth. Keep your own styles outside the scope and the two stop fighting over specificity.

Vercel folds Data Cache into Runtime Cache

Next.js fetch caching and unstable_cache now run on Vercel's Runtime Cache with no code changes, per Vercel's post. Pro storage becomes per project and environment at eight times the old team-wide pool, Enterprise at ten times.

The catch for client sites: Data Cache reads and writes that were never billed now follow Runtime Cache pricing. Check the usage tab on any marketing site that leans on cached fetches before the invoice does it for you.

Build a run-compare landing page

Build a single-page "run compare" product page in the style of an experiment tracker's compare view. Use it to launch any tool that produces several versions of one thing (AI image models, pricing plans, design variants).

Structure:
1. Sticky top app bar: product name left, three text links right.
2. Hero, two columns. Left: a ticket-stub tag (filled orange half with an ID, outlined half with date), a mono kicker, a heavy grotesque headline capped at 15ch, a two-sentence deck, one solid button and one text link. Right: a light "Shared input" panel that prints the single input once, with a small labelled list: Held constant, Varied, Picked.
3. Compare strip: three equal columns of square output images on dark cards. Under each image, a run ID, a filled model label in that column's series color, and three short labelled rows describing what differs.
4. A light "Trace" panel as the table of contents: one row per section with a run ID, section name, a horizontal span bar on a ten-step track, and a duration.
5. Content sections as light cards on the dark ground, each opened by a header strip with a mono run ID, a filled status label (Shipped, Recap, Changed) and a right-aligned source line, then a heavy h2.
6. One inverted dark card with a live "pick a run" radio group that dims the unpicked columns to 42 percent opacity and outlines the pick, using :has(), never removing them.
7. Sources, three pager tiles, footer.

Visual system: warm oxblood-black ground (#251214) with a faint two-scale plot grid (24px and 96px lines at 3 and 5.5 percent bone). Bone cards (#F2EADD) with near-black ink (#1E1112). Lantern orange (#E9893A) for buttons and run A; sage (#9DB8A2) and dusty rose (#D6A39C) for runs B and C. On light cards use darker siblings for text-weight color (#9E4316, #3F6650, #8A3F3A). Type: Pathway Extreme 800 for display, Petrona for body, IBM Plex Mono for IDs, labels and data.

Guardrails: body at least 19px with line height 1.65 to 1.75 and lines of 60 to 75 characters; WCAG AA for every label on its real background; cards 12px radius, labels 5px, no pills; one soft shadow on the shared-input panel only; hover and focus states at about 150ms with a prefers-reduced-motion guard; no neon, no glow, no purple-blue gradients, no gradient text; images with no legible text or logos; no fake search fields or toggles that do nothing.