A quiet experiment into prompting, empathy, and local-first design
What if AI could help someone practice everyday conversations, like ordering coffee or making a phone call, without feeling judged?
Not trying to fix anything, just encouraging. Not scoring or assessing, just guiding.
That question sparked a small pet project I’ve been working on: a lightweight, local-first app designed to support speech practice through gentle, AI-powered interaction. It’s not meant to replace therapy or formal support. Instead, it’s an exploration of what it might feel like to practice speaking in a space that feels safe, private, and kind, a space where encouragement takes priority over correction.
To my surprise, using the app felt remarkably similar to moments in real speech therapy: focused, reflective, and safe. While it’s no substitute for a human therapist, the AI’s calm, judgment-free responses created room to practice without pressure.
Whether it was suggesting simpler phrasing, offering pacing tips, or just responding with care, the experience reminded me that how we prompt and why matters just as much as what we build.
Prompting as emotional design: shaping AI to feel human
Generative AI is everywhere: writing, coding, creating art and music. It’s exciting, sometimes overwhelming, and undeniably powerful. These models can automate tasks, accelerate workflows, and unlock new forms of creativity.
But beyond the hype, something more subtle is emerging: the potential for AI to support personal growth, encourage reflection, and respond with empathy. The focus begins to shift, from what AI can do, to how we interact with it.
One of the biggest lessons from this project was realising that prompting isn’t just about asking questions. It’s a form of design.
In traditional apps, design lives in buttons, layouts, and user flows. With generative AI, the prompt becomes part of the design itself. It carries the voice, sets the tone, and shapes the emotional texture of the experience.
Prompts aren’t just technical inputs, they carry tone, intent, and emotional weight. A thoughtfully crafted prompt can transform a model’s response from mechanical to something that feels more like a guide, a coach, or even a companion. In that sense, prompting becomes a form of design—not just about functionality, but about shaping how the interaction feels.
That means writing prompts takes more than just technical skill. It needs:
-
- Empathy: anticipating how the user might feel, and using language that’s kind, respectful, and encouraging
- Clarity: being super clear about what you want the AI to do (and just as importantly, what you don’t)
- Cultural sensitivity: acknowledging differences in language, speech and background
- Lots of tweaking: trying things out, learning from what the AI says, and adjusting the prompts until they sound natural and helpful
Prompts are the bridge between what people want and what the AI says. And when they’re thoughtful prompts, that bridge doesn’t just transmit information; it carries tone, intent and care. That’s what makes the interaction feel less like a cold machine, and more like a caring companion.
Designing a gentle interaction flow
With that in mind, I thought about the feeling I wanted the app to create and shape the experience around that:
-
- Type a sentence you want to practice
- Get a few easier, friendlier ways to say it
- See gentle tips on pacing and word choice
- Chat with an AI that feels like a supportive conversation partner
- Reflect, without any pressure
And because it’s all local, your privacy stays safe.
Some prompt examples from the project
These principles came to life through a series of small, deliberate choices in the way I prompted the model.
- Rephrasings: Instead of simply telling the AI to “rewrite this sentence,” I asked it to offer three short, easy-to-say alternatives that avoid tricky sounds and complex phrasing. The prompt also reminded the AI to be kind and accessible, not perfect or rigid.
- Pacing advice: The AI gently suggested where to pause for breath, stretch vowels, or speak clearly. Its goal was not to be clinical, but calm and give practical guidance with a warm tone.
- Roleplay: The AI responded as a friendly conversation buddy: encouraging, informal and never too wordy. It wasn’t trying to fix; it was just there to practice with.
These small shifts in prompting made a big difference.They shaped the experience into something supportive, rather than corrective. Something you could return to without fear of judgement.
Designing for emotional safety
When you’re building something that touches on communication, especially the kind that can feel vulnerable, emotional safety becomes as important as functionality.
The goal of this app was never to evaluate or correct. It was to create a space where someone could practice without pressure.
That meant making deliberate choices in the user experience. The app avoids anything that might introduce anxiety or invite comparison:
-
- No correctness scores
- No tracking dashboards
- No backend logging
- No AI judgment
Just gentle suggestions, friendly roleplay, and calm space to reflect peacefully.
When images speak softly
Design doesn’t stop at words. Visuals carry emotional weight too, and they shape how an experience feels, even before a single sentence is read
To reflect the same tone of calm, care and inclusion, I used ChatGPT’s image generation to explore a different kind of visual language. The focus wasn’t slick UI mockups or polished graphics, it was emotional texture.
I prompted for themes like:
-
- Engaging with technology in harmony
- “Speech therapy across generations”
- “Vibrant microphones and speech pattern”
These AI-generated images replaced generic stock photos and helped create a visual tone that felt soft, warm and welcoming. More like an invitation than a user interface.
Working with local tools and closed loops
The app was built using Cursor, a conversational code editor, and Ollama running LLaMA3 locally. FastAPI handled the backend, and React the frontend. Each choice supported not just the functionality of the app, but the values behind it: privacy, responsiveness, and a deeply iterative, conversational way of working.
Why Cursor?
Cursor quickly became a natural choice because it treats coding like a conversation. Instead of switching constantly between code, terminal, and documentation, I could ask, test ideas, and refactor in-line, almost like pair programming with an assistant who understood the context.
For a project rooted in prompting, where small shifts in language can change the entire outcome, that ability to iterate quickly and conversationally made all the difference.
Why LLaMA 3 via Ollama?
LLaMA 3 struck the right balance between quality and performance, especially when fine-tuned or guided through well-crafted prompt engineering. Running it through Ollama offered a clean, developer-friendly way to manage and test models locally, crucial for keeping everything on-device.
This wasn’t just about performance or cost. It was about ownership: having a private, offline AI partner that respected the sensitivity of voice data and didn’t rely on cloud APIs or external tracking. It made testing feel safe and personal, and more like working with a local coach than a remote server.
Why FastAPI?
FastAPI provided the right mix of modern Python tooling, speed, and clarity. Its support for asynchronous requests and automatic documentation meant I could move quickly without sacrificing structure. For a lightweight app that needed to listen to user input, query the model, and respond cleanly, all while remaining easy to test, FastAPI hit the sweet spot. It also made local deployment frictionless and kept the project small, portable, and maintainable.
What this setup enabled
-
- Privacy first: Nothing leaves your machine
- Fast iteration: Try a prompt, tweak it, test it again
- Conversational debugging: When something goes wrong, you get a back-and-forth instead of a frustrating error
- Prompt and code shaping side by side: You can instantly see how prompt changes affect the app.
It made development feel like a real conversation between me and the AI.
What I learned along the way
This project deepened something I already sensed: prompting isn’t just technical, it’s emotional design. The way you ask shapes how users feel. A prompt can be cold and robotic, or kind and welcoming. That tone becomes the voice of the experience itself.
- Thoughtful prompts foster psychological safety.
- AI-generated images, when guided with care, invite warmth and inclusion.
- Local models protect privacy and create space to experiment without fear.
Together, these choices showed me that AI can be more than a tool. It can be a partner in supporting personal growth.
Final thoughts
This project is one of many quiet experiments happening at the edges of the AI boom. Its biggest value wasn’t in the feature set, but in the questions it raises and the perspective it offered.
As we continue building AI-powered tools, perhaps the most important thing won’t be what they do, but how we build them: with thoughtful prompting, human-machine collaboration, and a real desire to create safe, caring spaces.
Even a small, local prototype built with attention and intention can shine a light on how technology might help us grow with patience, kindness, and humanity.
