Taste rebuilds landing pages from screenshots with Codex or Claude Code

• Taste is being used as a taste layer for coding agents: MengTo's Taste post said the skill can pair with Codex or Claude Code to recreate landing pages from a video or full-page screenshot.
• Image generation moved upstream of implementation: Rohan Varma, a Codex PM, used Image Gen for 4 to 5 UI directions, then turned one image into a Codex Site in a Codex prototyping clip.
• Fable 5 pushed the workflow toward giant production prompts: MengTo's Koisei prompt specified stack, assets, shaders, scroll choreography, performance rules, and an acceptance checklist.
• Voila launched as a browser-side visual AI devtool: the Voila launch post said builders can point at UI and turn requests into code, prompts, copy, or tasks.

MengTo posted the Reddit thread that inspired Koisei, the shared result, and the full master prompt inside the tweet. Voila points to a browser tool, while Peter Yang's Codex conversation is available as a full episode. The common move is simple and useful: screenshot or screen-record the thing, generate the taste reference, then hand the agent a much more specific target.

Taste works like a style adapter for agents that already know how to write code. MengTo said the Taste link pairs with Codex or Claude Code to recreate almost any landing page from a video or full-page screenshot.

He later said the same approach works for image generation, brand direction, image-to-code, styles, and general design direction in a follow-up reply. The opinionated part is the product: MengTo contrasted it with default purple gradients in one reply and said the workflow still needs humans in another follow-up.

In one design reply, he reduced the shift to a designer joke: “we don't move rectangles anymore.”

Rohan Varma, a Codex PM, used image generation as the first design pass instead of starting in code. The workflow in Peter Yang's clip was:

1. Take a screenshot of the Codex composer.
2. Ask Codex to use Image Gen to prototype 4 to 5 ideas for project selection.
3. Pick the strongest generated image.
4. Ask Codex to turn that image into a working Codex Site.

Peter Yang's longer episode post framed Codex as Rohan's “everything app” for design prototypes, automations, live Codex Sites, and using one Codex thread to manage others. Yang quoted Rohan saying he was doing “3-4x more” over a day with Codex.

MengTo's Koisei prompt reads like a production brief for an interactive editorial site. In his workflow note, he said the build combined screen recording to prompt, video generation, and a detailed Fable 5 prompt that included every scroll interaction, the libraries to use, and the story flow section by section.

The prompt specified:

• Stack: Vanilla HTML/CSS/JS or Vite, three.js, GSAP ScrollTrigger, Lenis, scroll-scrubbed video, and no UI frameworks.
• Assets: exact image and video URLs for hero, koi river, temple canal, kimono lane, lantern water, night lantern, and two scroll films.
• Design system: named color tokens, display serif, UI sans, Japanese accents, film grain, mat borders, and soft paper vignettes.
• Motion: preloader, fixed nav, petal particle system, ripple shader, pinned hero, scroll films, horizontal gallery, day-to-night dissolve, custom cursor, and footer wordmark reveal.
• Performance: one shared WebGL renderer, devicePixelRatio capped at 1.75, offscreen pause with IntersectionObserver, metadata-first video loading, mobile pin reductions, and reduced-motion fallbacks.

According to the one-shot post, MengTo argued that with Fable 5 “the prompt is quickly becoming the most valuable asset.” In a reply about sharing it, he said the technique mattered more than copying the exact prompt.

The workflow around Fable 5 was prompt-first, then implementation. MengTo described the process as refining assets, source references, and prompt details until the prompt produces better results, rather than iterating the final page directly.

The mechanics were explicit:

• Fine-tune the prompt before fine-tuning the final output, per MengTo's process reply.
• Use one detailed prompt instead of a long agent rabbit hole, according to his cost reply and his Sonnet and Opus comparison.
• Keep the sequence as “brief -> prompt -> html -> iterations,” because MengTo's workflow reply said going to HTML too early costs more tokens.
• Audit claude.md, make skills less prescriptive, and include the “why” behind the request, per Allie Miller's Fable prompting tips.

Voila launched as a visual AI devtool for builders. The launch post described a browser workflow: point at any UI, ask for a change, and turn the selected interface into code, prompts, copy, or tasks.

The product's demo language matches the same screenshot-to-spec pattern as Taste and Fable, but aimed at live UI inspection. A later Voila demo reply showed output mode switching across code, copy, tasks, and prompts.

The launch was not a cold drop. In lucchaissac's thank-you post, lucchaissac credited a feedback group before posting the public link.

Voila can inspect local or live websites, but its edit model has boundaries. lucchaissac said changes are saved locally until reset, the extension does not inject JavaScript for security reasons, and saving back through PRs or agents is available on Enterprise.

Before the public launch, a Chrome Web Store clip said Voila was waiting for Chrome Web Store approval.