What Perplexity, Notion, and Anthropic Teach Us About AI UX
Nov 10, 2025
6 min read
1. First principle: get hands-on with AI to design it well
Designers build better AI products when they actively “shape the clay” of AI tools, not just read about them. Designers should accomplish prompting, adjusting, and studying how small design decisions for an AI product can affect user understanding. This hands-on work reveals that designing for AI is less about aesthetics and more about exposing reasoning, constraints, and feedback loops so that people can follow how the system thinks.
How We Keep AI Sensitivity in Our Design Practice
Work in live AI surfaces daily.
At VSDesign, we’ve learned that the only way to design for AI is to design with it. Our UX team works in live AI environments every day, testing and documenting how users actually interact with language-based systems. Whenever a result feels opaque or detached from the user’s intent, we record the moment and note what interface detail could make it clearer, whether that’s revealing how the model interpreted the prompt, clarifying the system’s goal, or surfacing the data behind it. Each observation is tagged with a pattern name so the whole team can reuse solutions instead of reinventing them. This ongoing log has become our shared design language, helping us quickly identify recurring issues and consistent fixes. Over time, it’s evolved into a living library of design knowledge that keeps our practice aligned and responsive to how AI truly behaves.
Evaluate with simple heuristics.
We also apply a set of simple heuristics in every prototype and product test. Each prototype or product test can begin with a few grounding questions: Did the interface make both user intent and system intent visible before running a prompt? Could users tell what changed between one version and the next? Did the product teach them something about how the AI works? These checkpoints keep our focus on clarity and user control, ensuring that innovation never comes at the cost of understanding.
Realizing the biggest blocker: moving users from clicking to writing
Most people learned software as a sequence of clicks and taps. AI tools shift the center of gravity to reading and writing. That leap creates anxiety: “What do I type? How does it work?” Perplexity’s product strategy shows practical ways to smooth that leap with answers like cite sources, progressive clarification, and prompt scaffolds near the input.
To see how thoughtful design can ease this transition, we can look at several examples, including Perplexity AI, Notion AI, Typeform. Their interface demonstrates practical ways to bridge the gap between traditional UI behavior and language-based interaction and help users understand what the system can do while keeping them in control. The result is less guesswork, clearer feedback, and a growing sense of trust in how the AI thinks.
a. Example of Perplexity - Answer with receipts
Put citations next to generated text so users can audit and learn. NN/g’s interview with Perplexity’s head of design highlights how clear sourcing builds trust and comprehension. In practice, this means the answer area includes a “Sources” section that lists the original web links the AI pulled from, allowing users to click and verify details directly. Each key point in the response is also marked with footnote-style numbers, connecting claims to their sources at a glance. When a user hovers over or taps a reference, the corresponding source preview appears, helping them decide whether to dive deeper. This simple structure turns an abstract, model-generated answer into a transparent dialogue grounded in evidence, showing users both what the AI says and where it comes from.
Pic from The NN/g UX Podcast: Perplexity AI’s answers provide the sources at the very top of each response.
b. Example of Perplexity – Guide intent with follow-ups.
Meet vague queries with precise follow-up questions and light structure, rather than a blank textbox. Reporting on Perplexity’s “answer engine” describes this path as moving from search to intention capture.
Pic from Emily Stevens- Perplexity AI‘’s precise follow-up questions
c. Example of Notion’s AI - Teach by example near the start.
Initial galleries, suggestions, and small prompt templates reduce “blank page” fear and show range before users write. Shape of AI’s Initial CTA, Gallery, and Suggestions patterns capture this.
Notion AI‘s complications emphasize additions to add context
AI is presented alongside typical options in Typeform's wizard
3. Turn prompting into product: designing a Prompt Enhancer users can trust
Bridging that language gap is only the first step. Once users feel confident enough to type their intent, the next challenge is helping them refine it. Many users know what they want conceptually, but struggle to express it in a way the model understands. They might over-explain, under-specify, or rely on everyday phrasing that the system interprets too literally. This is where design can actively support the translation between human thought and machine instruction. Instead of leaving users to guess what makes an “effective prompt,” we can build interfaces that guide, edit, and clarify those inputs in real time, which is turning prompting itself into part of the product experience.
Models already rewrite user input behind the scenes. Bringing that process into the open makes results better and computation cheaper. A Prompt Enhancer sits by the input, proposes concrete additions, and turns rough intent into a clear, constrained instruction before anything runs. Treat this as a first-class feature.
What your Prompt Enhancer should do:
A good Prompt Enhancer doesn’t hide behind automation. When users click “Improve” or “Refine,” the interface should display exactly what the system adds or changes. In practice, this can be done through a small toolbar action placed near the prompt input. Once triggered, the enhancement appears directly inside the text box, with subtle highlights indicating newly added constraints, tone adjustments, or clarifying phrases. The user can review, edit, or undo these changes before submitting.
This simple workflow turns improvement into a visible, collaborative step rather than a hidden rewrite. By keeping the interaction within the input field, users stay anchored in their own words while gaining the benefit of structured assistance. It mirrors the Prompt Details and Restructure patterns, making the invisible logic of AI editing
Offer small, legible boosts. Present bite-size chips like “add examples,” “set audience,” “limit word count,” “cite sources,” “include tone.” These map to Tuners and Prompt actions so users control context and style without learning prompt jargon.
UI patterns that reduce cognitive load
a. Inline enhancer with sparkles. Co-locate the enhancer next to the input and use clear iconography users already associate with “improve.” Shape of AI catalogs common visual language for enhance and summarize.
b. Describe ↔ Enhance loop. Pair Describe (turn output into transparent text) with Enhance so users can convert any result back into actionable prompts and iterate quickly.
c. Chained actions with guardrails. When the prompt clearly lacks context, show a one-tap path to add missing details or switch modes, rather than running something brittle.
After enhancement, show what changed, why it helps, and how to undo.
Perplexity’s public docs also model transparent system guidance: explain which instructions the system follows, where real-time search applies, and where it does not. Users understand limits and adapt faster.
At VSDesign, we’ve explored the AI UX transparency deeply in the other blog, showing how intentional UX systems can turn opacity into clarity and skepticism into confidence.
4. A simple AiUX workflow you can run this week
Step 1. Map the jobs. List top tasks where users feel stuck writing. Validate by watching queries in your logs or running 5 short usability sessions.
Step 2. Seed patterns. Add Initial CTA with 4-6 suggestions and a tiny gallery of good prompts plus visible prompt details.
Step 3. Ship a minimal Enhancer. Start with three tuners: audience, constraints, and references. Show a diff view of the rewritten prompt. Let users toggle each tuner.
Step 4. Add a sample run. Preview format with one record or a small excerpt before full execution.
Step 5. Measure what matters. Track reduction in abandoned prompts, increased first-run success, fewer re-prompts, and fewer support tickets about “what should I type.”
Conclusion
AI products are moving from navigation to intention. Users no longer want to click through layers of menus; they expect direct, transparent answers that explain themselves. Perplexity’s evolution captures this shift: people now look for tools that let them express what they mean without learning the hidden rules of prompting. For designers, the challenge is to make language-based interaction as effortless and intuitive as a button click. Patterns like Prompt Enhancer and Prompt Details aren’t just design features; they are the bridge between human intention and machine reasoning.
At VSDesign, we help AI startups build that bridge. Our team combines UX strategy, product design, and AI interaction research to translate complex model behavior into clear, trustworthy user experiences. Whether you’re designing your first AI-powered MVP or refining an existing product, we help you prototype, test, and scale with clarity.
If you’re exploring how to make your AI product more intuitive and human-centered, let’s start a conversation.