PromptCheck
The Product
PromptCheck is a tool that reviews AI prompts for communication issues before they're sent — like spellcheck, but for how you talk to AI. It identifies vagueness, conflicting instructions, missing context, leading bias, and other patterns that lead to poor AI responses and wasted effort.
Timeline: May 2026
Platform: Browser Extension / Web Tool
Role: Sr. Product Designer
Tools: Claude, Claude Design, Claude Code
Responsibilities: Problem definition, conversational research, rapid wireframing, pattern exploration, prototype iteration, handoff to code
The Problem
Today's AI chat interfaces have a communication gap. Users tend to assume models can do more or infer more than they actually can, and write prompts that are vague, contradictory, or missing context. When the result isn't what they wanted, they re-prompt and try again. The model meanwhile fills in the gaps silently, without telling the user where its guesses were.
The Goal
Design a lightweight tool that flags prompt issues inline before the prompt is sent. The goal is to help users write clearer prompts without needing to know prompt engineering, and cut down on the back-and-forth re-prompting that wastes time and compute.
Defining the Problem Space
Conversational Research with Claude
Rather than starting with traditional user interviews, the first step was a structured conversation with Claude about the problem space. This dialogue mapped out a taxonomy of prompt communication issues — specific, repeatable patterns where the way a prompt is written leads to suboptimal AI output. The conversation served as both a research method and a design brief.
Mapping Prompt Issues
This research led to building a mockup that catalogued real examples of each issue type — color-coded by category, with annotated prompts showing exactly where each issue appears and why it matters. This artifact became the foundation for the design work that followed.
Issue taxonomy mockup with annotated example prompts
Issue Categories
The research identified two families of prompt issues: patterns that skew the answer (bias) and patterns that degrade the answer (quality).
Vagueness
Ambiguous terms like "a few," "recently," or "make it pop" that leave interpretation entirely to the model
Conflicting Instructions
Contradictory demands like "concise but thorough" or "formal but conversational" that force the model to pick a middle ground
Missing Context
References to "the last one" or "what we discussed" — information the model doesn't have access to
Leading Bias
Framing like "why is X destroying Y" that treats a claim as settled, pushing the model to validate rather than evaluate
Negation Traps
Telling the model what NOT to do — "don't be boring" seeds the concept of boring rather than describing the positive target
Missing Audience / Format
No audience or output format specified — "explain quantum entanglement" to a physicist vs. a child are entirely different responses
Exploration & Decision
Rapid Wireframing with Claude Design
Instead of sketching on paper or building static mockups in Figma, I used Claude Design to rapidly generate wireframe variants on a single canvas. Each variant explored a different interaction pattern for surfacing prompt issues. The speed of this process — generating a full exploration canvas in minutes rather than hours — made it possible to explore more of the solution space before committing to a direction.
~8 wireframe variants explored on a single Claude Design canvas
Variants Explored
Each variant represented a different trade-off between visibility, intrusiveness, and information density. One example prompt was threaded through all variants for apples-to-apples comparison.
Inline + Hover Tooltip
Issues are underlined directly in the prompt text. Hovering reveals a tooltip with the issue explanation and suggested fixes. Minimal visual disruption — the user stays in their writing flow.
Stacked Cards
All issues listed as cards below the prompt. Clear and scannable, but separates the issue from its location in the text.
Margin Gutter + Minimap
Issue indicators in a gutter alongside the prompt, like code editor lint markers. Familiar for developers, less intuitive for general users.
Focus / Review Mode
A dedicated review screen entered after writing. Thorough, but adds friction — users must context-switch from writing to reviewing.
Browser Extension Popover
A floating popover anchored to any AI chat composer. Platform-agnostic, but disconnected from the prompt text itself.
Choosing the Inline + Hover Tooltip
I chose the inline tooltip pattern because it does three things well at the same time:
Non-Disruptive
The user stays in their writing flow. Issues are surfaced contextually rather than in a separate view or modal.
Progressive Disclosure
Underlines provide a lightweight signal. The full explanation and fix options only appear on hover or click — detail on demand.
Spatial Context
Each issue is anchored to the exact text that triggered it, making the feedback immediately actionable.
Prototype & Handoff
From Claude Design to Claude Code
Once I'd settled on the inline tooltip pattern in Claude Design, I handed it to Claude Code to build out as a working HTML/CSS/JS prototype. There were no spec documents or design tokens to write up — the handoff happened in a single conversation.
Handoff from Claude Design to Claude Code
Working Prototype
The prototype demonstrates the full interaction flow: the user writes a prompt, clicks PromptCheck, issues are detected and underlined in-place, and hovering any mark reveals a tooltip with the issue explanation and quick-fix options. Applying a fix updates the prompt text in real time.
Interactive prototype — open full screen
Future Resolution Patterns
Beyond the inline tooltip quick-fixes, I explored three more resolution patterns in Claude Design as possible future features. Each gives the user a different level of control over how an issue gets fixed.
Conversational Fix
Guided questions that help the user clarify their intent before rewriting
Diff Preview
Side-by-side view of original vs. rewritten prompt with changes highlighted
Three Rewrites
Pick from alternative phrasings — tight, balanced, or deep dive
A New Design Workflow
I worked entirely with AI tools on this project — no Figma, no spec handoff.
1. Conversation
I started with a structured dialogue with Claude to define the problem space, map issue categories, and generate example prompts. For an early-stage project, the conversation took the place of desk research and interviews.
2. Claude Design
I used Claude Design to explore around eight wireframe variants on a single canvas. Having them side by side made it easier to weigh trade-offs and pick a direction in one session.
3. Claude Code
I handed the refined design to Claude Code for a working prototype. No specs, no design tokens, no redlines — the conversation was the brief.
Where the time goes now
With tools like Claude Code, the speed of engineering has increased so dramatically that implementation is no longer the bottleneck. When a working prototype can be generated in minutes, the constraint moves upstream — to ideas, solutions, and talking to users for feedback.
Tools like Claude Design keep design aligned with this new pace. Prototypes are concrete, can be created rapidly, and can be interacted with almost immediately — by other designers, developers, and users. They reduce coordination overhead and let everyone move quickly.
The design loop to compress: talk to users → design features → ship code → read feedback → repeat. The faster each step gets, the more cycles you can run — and the better the product becomes.
Conclusion
Result
PromptCheck went from problem definition to a working interactive prototype in a single design session. The inline tooltip pattern provides contextual feedback that helps users write better prompts — without requiring them to learn prompt engineering.
What I Learned
The biggest shift wasn't any one tool — it was the pace of iteration. When prototyping is nearly instant, the most valuable design skill becomes deciding which ideas to explore and when to move to user feedback.
Next Steps
Conduct user testing with real prompt-writing sessions to validate whether inline detection improves prompt quality
Test resolution patterns (conversational fix vs. diff vs. three rewrites) to determine which users prefer
Explore integration points — browser extension, chat UI plugin, API service
Investigate running the pre-check with a local model to further reduce data center usage
Thank you
Thank you for your time reviewing my work on PromptCheck!
If you'd like to get in touch, please say hi!
claytonanderson.work@gmail.com