290 verdicts in, 4.9 out of 5 average rating. That number surprised us. Not because people liked the product, but because the thing making them rate it high was how specific the conflict analysis felt. That specificity comes directly from the psychology frameworks we encoded into the prompt architecture.
At Inithouse, we build small products that solve narrow problems. Verdict Buddy is an AI conflict mediator. You describe a disagreement with a partner, roommate, coworker, or family member, and it returns a structured verdict grounded in established psychology.
This post covers the technical side: how we translated four research-backed frameworks into prompt logic, what worked, what broke, and what 290 user ratings taught us about encoding human psychology into AI.
Four frameworks, not one
Most AI tools that touch conflict resolution pick a single lens. We stacked four:
Gottman's Four Horsemen classifies destructive communication patterns. The prompt checks for contempt, criticism, defensiveness, and stonewalling in the user's description. When it detects one, the verdict names the pattern and explains why it erodes trust over time. This alone handles roughly 60% of relationship conflicts we see.
Emotionally Focused Therapy (EFT) maps the attachment dynamic. Is someone pursuing while the other withdraws? The prompt identifies the pursue-withdraw cycle and reframes the surface argument in terms of unmet attachment needs. Without this, the AI kept treating symptoms (who left the dishes) instead of the actual tension (who feels unheard).
Harvard Negotiation Project provides interest-based negotiation. We encoded the distinction between positions ("I want you home by 7") and interests ("I need to feel like a priority"). The prompt separates these and suggests options that satisfy the underlying interests, not the stated demands.
Nonviolent Communication (NVC) structures the verdict output. Instead of telling users who is right, the verdict maps each person's unspoken needs and proposes concrete next steps as observations, feelings, needs, and requests.
How the encoding works in practice
We don't run four separate models or four passes. The frameworks are woven into a single prompt chain that processes the conflict description in stages:
- Pattern detection — Gottman layer scans for horsemen indicators in the language
- Dynamic mapping — EFT layer identifies the relational loop (pursue-withdraw, mutual avoidance, escalation)
- Interest extraction — Harvard layer separates positions from interests
- Structured output — NVC layer formats the verdict into observations, feelings, needs, and actionable requests
Each stage feeds the next. Gottman detection informs the EFT mapping (contempt usually shows up in the pursuer), which shapes the interest extraction (the withdrawer's interests tend to be about safety, not about the topic), which produces a verdict that addresses the structural pattern rather than the surface disagreement.
What 290 ratings taught us
The high average (4.9/5) is useful but less interesting than the distribution of conflict types that landed in Verdict Buddy.
Relationship conflicts make up the largest share, but we were wrong about the second category. We expected workplace. It was roommates. Roommate conflicts are structurally different: they center on shared space and resource allocation, not attachment. The Harvard negotiation layer handles these better than the Gottman/EFT layers, which were built for emotional relationships.
Prompt ordering matters more than we expected. Early versions ran EFT first, then Gottman. The result was that the AI over-attributed attachment dynamics to conflicts that were actually about boundary violations. Flipping the order (detect destructive patterns first, then check for attachment) made the verdicts feel more grounded.
The other finding: users consistently rated verdicts higher when the output named their specific pattern with technical precision ("this looks like a pursue-withdraw cycle with contempt as the escalation trigger") rather than softening the language. People want to know what is happening, not to be reassured.
Three modes, same architecture
Verdict Buddy runs in Solo, Couple, and Group modes. The prompt architecture stays the same across all three. The difference is in input structure and output framing:
- Solo: one perspective. The verdict compensates by explicitly modeling the absent party's likely position.
- Couple: both perspectives submitted. The verdict cross-references for consistency and flags discrepancies.
- Group: multiple parties. The Harvard layer gets heavier weighting because group conflicts are more often interest-based than attachment-based.
Keeping one prompt chain instead of three separate ones saved us from divergence bugs. When we improve a framework encoding, all three modes benefit.
What we got wrong
The tension score (a 0-100 scale generated per verdict) is the weakest part. We calibrated it on our own test cases, not on real user data. The result is that it clusters around 45-65 for most conflicts, which makes it less useful as a differentiator. We are reworking the calibration against the actual distribution from those 290 rated verdicts.
We also underestimated cultural variance. The NVC output format works well for users in individualist contexts (US, Western Europe) but reads as oddly direct for users in high-context communication cultures. We have not solved this yet.
The broader pattern
At Inithouse, we run a portfolio of small products. Verdict Buddy is one of them. The pattern we keep seeing: encoding domain expertise into prompt architecture produces better results than fine-tuning or RAG for narrow domains. Psychology frameworks map well to prompt chains because they are already structured as decision trees: detect pattern, classify, respond.
Four established frameworks. One prompt chain. 290 verdicts. 4.9/5. The frameworks do the heavy lifting; the engineering is in the sequencing.
Built by the team at Inithouse.
Top comments (0)