Prompt Sensitivity

Prompt sensitivity is the extreme reactivity found in Google Stitch where changing a handful of words in your prompt produces dramatically different Figma-ready layouts. The tool does not gently evolve your idea. It leaps to new conclusions about hierarchy, spacing, component arrangement, and even which parts deserve auto layout. Launched in 2025 alongside Material 3 updates, Stitch inherited this trait from Gemini. The model treats every noun as a component request and every adjective as a layout directive. This makes it powerful for rapid exploration but demanding. The article highlights how the key elements section in the prompt template drives the biggest shifts. Name UserListWithAvatars and the output contains stacked circles with proper constraints. Forget to name it and that element disappears or appears as an afterthought. Sensitivity like this separates tools that feel like collaborators from tools that feel like unpredictable interns. By 2026 every serious prompt-to-UI system has it. Stitch just wears it on its sleeve more than most.

What it is not. It is not simply writing bad prompts although bad prompts make the problem worse. A vague prompt in a high sensitivity tool creates more dramatic failure than the same prompt in a more forgiving one like later versions of Uizard. It is not inconsistency across sessions. That is prompt drift and it is a different headache that hit early Gemini releases in 2024. Prompt sensitivity is immediate and reproducible. Run the same weak prompt twice and you get similar levels of garbage. Tweak three words toward specificity and both runs improve in similar ways. It is not a flaw that will be engineered away soon. The underlying architecture of these models makes sensitivity inevitable. The designers at Google who built Stitch understood this. That is why the export path emphasizes structure over surface polish. They want you to use the sensitivity for layout logic and then fix the visuals yourself in Figma.

Concrete example. Consider the team settings screen used as the example in the Stitch article. Weak version: Settings page for managing users. Stitch generated a single column mess with unlabeled buttons and a table that had no hierarchy between admin controls and regular user list. The avatars were missing. The role selector was a plain text field. The whole thing required 40 minutes of rework to reach usable. Strong version following the template: Team settings screen: Admin is adding a new member to the workspace. Layout: Two-column, sidebar nav on the left. Key elements: User list with avatars, invite form, role selector, permissions table. Tone: Dense but clear, no marketing copy. Constraints: Responsive, light mode, Material 3 tokens.

This time the output arrived with a left sidebar containing the nav, a main area with the user list showing circular avatars in a proper stack, an invite form with clear form fields, a dropdown role selector using M3 styles, and a full permissions table with toggle switches. Layer names were logical like userListContainer and permissionsTable. Auto layout was applied to the two column wrapper. The entire file imported cleanly and needed only 12 minutes of token remapping and surface updates. The gap is pure prompt sensitivity.

I saw identical behavior when testing Figma Make in March 2026. A loose prompt for a notification center produced a basic list. Adding specificity about card based layout with preview snippets and action buttons created a much more sophisticated component set. In v0 the sensitivity affects the code. One prompt got me a basic Tailwind div structure. Adding the exact component names from shadcn turned it into a composed set of reusable parts with proper variants. These examples show why the article stresses the template. The model is listening closely. Most designers are not speaking clearly enough.

A third example comes from dashboard work. Prompting for an analytics view with no details gave me three cards in a row and a chart placeholder. Specifying key elements like conversionFunnelChart, cohortRetentionTable, and realtimeMetricCards with dataHeavy tone produced a sophisticated layout with sidebar filters that matched production apps like PostHog from the article images. The grid logic was sound. The hierarchy worked at a glance. This is where sensitivity becomes an advantage instead of a bug. The same pattern destroyed early Galileo users in 2023 who kept feeding it marketing speak and wondered why every output looked like a generic landing page.

When to use and when not to. Account for prompt sensitivity in the three workflows where Stitch shines according to the article. During early exploration it lets you spin up five layout directions quickly by varying one parameter per generation. For layout drafts of unfamiliar patterns like nested permission screens or dense configuration pages it provides structurally sound bones that you can skin later. For Figma handoff preparation the sensitivity helps you get named layers and grouped elements that survive the round trip. Use it when you have 25 minutes to invest in a better starting point than a blank frame. Build and save prompt templates that you refine over time. Test them on low stakes screens first. Treat the template like a brief you would give a junior designer who interprets everything literally.

Stop using it when you reach surface design. Stitch and tools like it still produce mediocre typography and color choices in 2026. Sensitivity to tone descriptors helps only marginally here. Do not use it when you need interactive states because the tool only outputs static frames. The sensitivity cannot help you with hover states or loading animations. Avoid depending on it for client deliverables while it remains in Google Labs as the article warns. The product could change without notice. Do not build team processes that assume stable behavior. If your work is code first then v0 sensitivity is what you should master instead. Different tool, similar principle but applied to React components rather than Figma layers.

Prompt sensitivity is the reason good prompt templates feel like cheat codes in 2026 design tools.