We're building Verdict Buddy — an AI-powered conflict resolution tool that translates established psychology frameworks into structured prompts. Here's the technical story of how we got there and why prompt engineering for empathy turned out to be the hardest challenge we've tackled at Inithouse.
The Problem With "Just Talk It Out"
Most conflict resolution advice boils down to "communicate better." But when you're in the middle of a heated argument — whether it's with your partner about finances, your roommate about dishes, or your coworker about project scope — your brain isn't exactly in problem-solving mode.
We wanted to build something that could actually guide people through structured conflict resolution, the same way a therapist or mediator would. Not by replacing human connection, but by providing the framework when emotions are running high and you can't remember what your couples therapist said last Tuesday.
Four Frameworks, Three Modes
We studied four established approaches to conflict resolution and translated each into prompt architectures:
Gottman Method — focuses on "soft startups" and repair attempts. The prompt structure here emphasizes identifying the underlying need behind the complaint. Instead of "you never clean up," the AI guides toward "I need to feel like we're a team."
Emotionally Focused Therapy (EFT) — works with attachment patterns. We built prompts that identify whether someone is pursuing or withdrawing, then scaffold responses that address the attachment need rather than the surface issue.
Harvard Negotiation Project — separates positions from interests. The prompt chain here is almost algorithmic: identify stated positions, dig for underlying interests, generate options for mutual gain, apply objective criteria.
Nonviolent Communication (NVC) — observation, feeling, need, request. This was actually the easiest to implement as prompts because it's already quite structured, but the hardest to make sound natural rather than robotic.
Each framework maps to different conflict types. Gottman and EFT work best for intimate relationships. Harvard Negotiation is ideal for workplace or business disputes. NVC is the Swiss Army knife that works everywhere but requires the most prompt finessing to avoid sounding like a therapy textbook.
The Three Modes
Verdict Buddy operates in three distinct modes:
Solo Mode — you describe the conflict from your perspective. The AI asks clarifying questions using the selected framework, then provides insights about what might be driving the other person's behavior and suggests specific language for your next conversation.
Couple Mode — both parties describe their perspective through a shared link. The AI synthesizes both views, identifies where the real disconnect is (often it's not what either person thinks), and proposes a path forward that addresses both people's underlying needs.
Group Mode — for team conflicts, family disputes, or roommate situations with more than two parties. The prompt architecture here is significantly more complex because you're tracking multiple attachment styles, interests, and communication patterns simultaneously.
Why Empathy Is Harder Than Image Generation
Here's something that surprised us: getting an LLM to generate empathetic, framework-appropriate responses is significantly harder than most other prompt engineering tasks.
The failure modes are subtle. An image generator either produces something that looks right or doesn't. But a conflict resolution response can sound reasonable while actually being harmful — validating one person's perspective in a way that dismisses the other, or using NVC language in a way that comes across as passive-aggressive rather than connecting.
We spent weeks on what we call "empathy calibration" — making sure the AI doesn't:
- Take sides while appearing neutral
- Use therapeutic language that feels condescending
- Suggest solutions before both parties feel heard
- Default to compromise when the real issue is a boundary violation
The prompt engineering approach that finally worked was layering: the base prompt sets the framework constraints, a second layer adds the relational context (are these romantic partners? colleagues? family?), and a third layer handles tone calibration based on the emotional temperature of the conversation.
What We Learned About Prompt Architecture
Building this taught us something about AI product development that I haven't seen discussed much: the hardest prompt engineering problems aren't about getting accurate outputs — they're about getting emotionally appropriate ones.
You can't unit test empathy the way you test factual accuracy. We ended up building our own evaluation rubric based on therapy research about what makes interventions effective vs. harmful. Each response gets scored on: validation (does both parties feel heard?), reframing (does it shift from blame to understanding?), actionability (is there a clear next step?), and safety (does it avoid reinforcing harmful dynamics?).
Current State
Verdict Buddy is live and we're actively collecting feedback on how people use the different modes and frameworks. Early signals suggest that Solo Mode is the most-used feature — people want to understand their own conflicts better before engaging the other party.
If you're interested in the intersection of psychology and AI, or if you're building products that need emotionally intelligent outputs, I'd love to hear how you're approaching similar challenges.
I'm Jakub, building a portfolio of AI-powered products at Inithouse. Verdict Buddy is one of about a dozen MVPs we're running simultaneously to find product-market fit.
