Work
About
Resume
Playground

Back

Summary

Research

Design Goals

Design & Exploration

Final Design

Reflection

AI Image Generator for Educators

The project

An AI image generator built for K-12 and higher-education teachers — creating classroom-ready visuals by anticipating educator needs, so they can spend less time prompting and more time teaching.

My role

Product designer on a cross-functional team, translating PM research into UX flows, interaction design, and final UI.

Project duration

5 weeks

Tools

Figma

Notion

Why this matters

Teachers are losing time to the very tools designed to save it.

Most educators aren't AI prompt experts, and generic image generators weren't build with them in mind, causing teachers to spend more time correcting outputs than creating lesson materials. Research surfaced a compounding problem: the more teachers try to refine a prompt, the further they get from their original goal.

What educators need

Speed without sacrifice: Teachers need to produce usable, age-appropriate visuals quickly without compromising on content safety or relevance to their curriculum.

A tool that meets them halfway: Rather than expecting teachers to learn prompt engineering, the tool needed to anticipate their intent and guide them toward a good result.

The solution

An anticipatory image generator that turns Subject, Grade, and Learning Topic into smart suggestions so teachers get the right image in fewer prompts.

Research & Strategy

Stakeholder interviews & literature review

To build context before designing, I met with our PM to understand the project goals, user constraints, and the problem space he'd identified. From there, I conducted my own literature review to ground the problem in external evidence.

Research from the National Centre for AI revealed that educators were expected to write like photographers and think in "visually rich prompts", a skillset most teachers simply don't have and shouldn't need to develop. A study of 57 educators across Sweden and Australia reinforced this, finding that teachers frequently ended up reworking AI outputs themselves, adding time rather than saving it.

To build context before designing, I met with our PM to understand the project goals, user constraints, and the problem space he'd identified. From there, I conducted my own literature review to ground the problem in external evidence.

Research from the National Centre for AI revealed that educators were expected to write like photographers and think in "visually rich prompts", a skillset most teachers simply don't have and shouldn't need to develop. A study of 57 educators across Sweden and Australia reinforced this, finding that teachers frequently ended up reworking AI outputs themselves, adding time rather than saving it.

"I’m actually using more time refining the prompt than I am if I just did it off the top of my head. It’s a false economy, isn’t it"
– Legal studies educator (AU)

Competitive analysis

I analyzed four AI image tools educators commonly reach for, evaluating each for prompt guidance, output quality, style control, and classroom sustainability.

Each tool had something to offer, but none of them were built to anticipate what an educator actually needed next. That gap became the foundation for the design.

Emerging themes

Lower the floor for non-technical users: Teachers shouldn't need prompt expertise to get a usable image. The tool needs to meet them at their actual skill level.

Reduce time lost to trial and error: Every existing tool put the burden of refinement entirely on the user. The goal was to make the tool do more work so teachers could spend more time engaging with students.

Make classroom-readiness the default: Rather than generating generic images that teachers had to evaluate for appropriateness, the tool should treat grade level, subject, and content safety as foundational inputs.

Design & Exploration

Iteration 1: Simplifying the Input Flow

Problem

Competitive analysis revealed that every existing tool dropped educators into a blank text field with no context or guidance. The assumption was that users already knew what they wanted and how to describe it, but that's not always true for non-technical users.

Hypothesis

If I could capture a few key pieces of educator context upfront (grade level, subject, and image type) the tool could use those inputs to narrow the generation space and reduce the guesswork on both sides.

Hypothesis

If I could capture a few key pieces of educator context upfront (grade level, subject, and image type) the tool could use those inputs to narrow the generation space and reduce the guesswork on both sides.

I drafted a structured onboarding flow using card-based selections to make the inputs feel approachable rather than technical. The goal was to get teachers to a contextualized starting point without typing a single word.

  • Step 1: User indicates grade level and subject/learning goal.

  • Step 2: User select the type of image they want to create.

  • Step 3: User chooses another card to indicate image style.

  • Step 4: Users interact with the AI agent via text-based inputs.

What I learned

When I reviewed the designs with my PM, he flagged that while the card-based setup flow was structured and approachable, it left teachers with no clear path forward once results didn't match expectations. If educators couldn't refine outputs from within the flow they would have to abandon their selections and start over, compounding the exact fatigue the design was meant to reduce.

When I reviewed the designs with my PM, he flagged that while the card-based setup flow was structured and approachable, it left teachers with no clear path forward once results didn't match expectations. If educators couldn't refine outputs from within the flow they would have to abandon their selections and start over, compounding the exact fatigue the design was meant to reduce.

Iteration 2: Adding Anticipatory Suggestions

Problem

If teachers received a result they weren't happy with, the path forward was unclear. Without guidance, most would either retype their original prompt with minor variations or abandon the session entirely.

Hypothesis

If the system could anticipate likely next steps and surface them as one-tap options, teachers wouldn't need to know what to ask for next. The goal is to make the tool could do that thinking for them.

I designed a conversational flow where the AI surfaced contextual refinement suggestions after each generation. If a suggestion was rejected, the system offered alternatives rather than leaving the user at a dead end.

  • Initial Inputs: Teachers start by entering grade level and subject, giving the AI a strong foundation for tailored results.

  • First round of suggestions: The AI suggests possible prompt refinements right away, helping teachers move forward faster without needed to rewrite prompts.

  • Style selection for consistency: Teachers choose a style once and apply it across materials, ensuring visual consistency throughout their lesson content.

  • Confirmation setup: Once a suggestion is accepted, the AI confirms choices to ensure accuracy before generating the image, reducing misaligned outputs.

  • Start of image generation: After inputs are confirmed the AI begins producing an image, giving teachers a clearer sense of progress in the flow.

  • Initial Inputs: Teachers start by entering grade level and subject, giving the AI a strong foundation for tailored results.

  • First round of suggestions: The AI suggests possible prompt refinements right away, helping teachers move forward faster without needed to rewrite prompts.

  • Expanded guidance: If the teachers reject a suggestion, the system provides additional options, preventing dead ends in the workflow.

  • Confirmation setup: Once a suggestion is accepted, the AI confirms choices to ensure accuracy before generating the image, reducing misaligned outputs.

  • Start of image generation: After inputs are confirmed the AI begins producing an image, giving teachers a clearer sense of progress in the flow.

  • Initial Inputs: Teachers start by entering grade level and subject, giving the AI a strong foundation for tailored results.

  • First round of suggestions: The AI suggests possible prompt refinements right away, helping teachers move forward faster without needing to rewrite prompts.

  • Expanded guidance: If the teachers reject a suggestion, the system provides additional options, preventing dead ends in the workflow.

  • Confirmation setup: Once a suggestion is accepted, the AI confirms choices to ensure accuracy before generating the image, reducing misaligned outputs.

  • Start of image generation: After inputs are confirmed the AI begins producing an image, giving teachers a clearer sense of progress in the flow.

What I learned

Reviewing with my PM surfaced a new issue: splitting the experience across multiple steps was creating its own friction. The setup and suggestion layers felt disconnected, and the text input field was being underused. The flow needed to be consolidated so that guidance and input lived in the same place rather than across separate screens.

Reviewing with my PM surfaced a new issue: splitting the experience across multiple steps was creating its own friction. The setup and suggestion layers felt disconnected, and the text input field was being underused. The flow needed to be consolidated so that guidance and input lived in the same place rather than across separate screens.

Final Design

The iteration process taught me two things: structured inputs reduced cognitive load at the start and anticipatory suggestions reduced it during refinement. The final design brought both together into a single, unified experience.

The core change

The final design front-loads Grade, Subject, and Learning Topic directly into the main interface, giving the AI what it needs to generate relevant suggestions from the first interaction. Additionally, the text input field, previously underused, becomes the active surface where suggestions appear, selections are confirmed, and refinements happen.

  • Front-loaded inputs: Teachers provide key information upfront, reducing friction later and setting the AI up for more accurate suggestions.

  • Smart image suggestions: The system anticipates needs by suggesting classroom-friendly image types (e.g., diagram, chart, illustration), cutting down trial-and-error prompt crafting.

  • Style selection for consistency: Teachers choose a style once and apply it across materials, ensuring visual consistency throughout their lesson content.

  • Flexible text input: The text box is reframed as an optional detail-attention tool, giving educators flexibility without requiring heavy typing.

  • Initial image preview: AI generates a few image variations that the user can choose to refine further.

  • One-tap refinement options: AI generates an initial image and offers suggested refinements, reducing cognitive load and helping teachers get to a usable result faster.

  • Iterative refinements: Teachers can continue to refine and polish their images through guided options or move on to finalizing and downloading assets, avoiding the fatigue of re-prompting from scratch.

  • Export + save for lesson use: Final images can be downloaded, added to a lesson plan, or stored in a library, directly supporting classroom workflows.

  • Front-loaded inputs: Teachers provide key information upfront, reducing friction later and setting the AI up for more accurate suggestions.

  • Smart image suggestions: The system anticipates needs by suggesting classroom-friendly image types (diagram, chart, illustration), cutting down trial-and-error prompt crafting.

  • Style selection for consistency: Teachers choose a style once and apply it across materials, ensuring visual consistency throughout their lesson content.

  • Flexible text input: The text box is reframed as an optional detail-attention tool, giving educators flexibility without requiring heavy typing.

  • First image preview + refinement options: AI generates an initial image and offers one-tap refinements, reducing retries and helping teachers get to a usable result faster.

  • Iterative refinements: Teachers can continue to refine and polish their images with guided options or finish generating, avoiding the fatigue of re-prompting from scratch.

  • Export + save for lesson use: Final images can be downloaded, added to a lesson plan, or stored in a library, directly supporting classroom workflows.

  • Front-loaded inputs: Teachers provide key information upfront, reducing friction later and setting the AI up for more accurate suggestions.

  • Smart image suggestions: The system anticipates needs by suggesting classroom-friendly image types (e.g., diagram, chart, illustration), cutting down trial-and-error prompt crafting.

  • Style selection for consistency: Teachers choose a style once and apply it across materials, ensuring visual consistency throughout their lesson content.

  • Flexible text input: The text box is reframed as an optional detail-attention tool, giving educators flexibility without requiring heavy typing.

  • Initial image preview: AI generates a few image variations that the user can choose to refine further.

  • One-tap refinement options: AI generates an initial image and offers suggested refinements, reducing cognitive load and helping teachers get to a usable result faster.

  • Iterative refinements: Teachers can continue to refine and polish their images through guided options or finalize and download assets, avoiding the fatigue of re-prompting from scratch.

  • Export + save for lesson use: Final images can be downloaded, added to a lesson plan, or stored in a library, directly supporting classroom workflows.

What this eliminated

Teachers no longer face a blank prompt field. They no longer restart from scratch when a result misses. And they no longer move through multiple screens to accomplish what should be a single, fluid task.

Reflection

This was my first project working within an AI product context, and the biggest shift it required wasn't visual, it was conceptual. Designing for a generative interface meant accounting for uncertainty (i.e., "What happens when the output misses?", "What does the teacher do next?"). The edge cases that emerged throughout this project drove more of my decisions than the happy path did.

While the project didn't reach launch due to organizational changes at Reality AI, the value of this work sits in the process. Working from a PM brief, conducting my own literature review, running competitive analysis, and iterating through stakeholder feedback gave me a structured foundation for tackling AI-native design problems that I'll continue to carry into any future project in this space.

The natural next step would be user testing, specifically whether the suggestion chips reduce prompting burden or introduce a different kind of decision fatigue. I'd also want to explore how the tool performs across grade levels, since a kindergarten teacher and a high school teacher have very different needs.

Other Projects

Other Projects