Turns out the hardest part of building an AI conflict resolver wasn't the AI. It was figuring out which psychology framework to apply — and when.
I'm building VerdictBuddy, a tool that helps people navigate relationship conflicts using AI-powered analysis grounded in real psychological frameworks. Not generic "communicate better" advice — structured verdicts based on Gottman, NVC, attachment theory, and cognitive-behavioral models.
Here's the technical breakdown of how I got there.
The core insight: frameworks are prompt templates
Every established psychology framework follows a predictable structure. Gottman's Four Horsemen model, for example, identifies four destructive communication patterns (criticism, contempt, stonewalling, defensiveness) and maps each to a repair strategy.
That structure translates almost directly into a prompt template:
- Input classification — which pattern is present in the user's conflict description
- Framework logic — what does the theory say about this specific pattern
- Output generation — concrete verdict + reasoning + actionable next steps
The same applies to Nonviolent Communication (NVC), which breaks every conflict into observations, feelings, needs, and requests. That four-part structure becomes a four-stage prompt pipeline.
Framework selection logic
Not every framework fits every conflict. A workplace disagreement about project ownership doesn't need attachment theory. A recurring argument with a partner probably doesn't need a pure CBT reframe.
I built a classification layer that analyzes the conflict description and selects the most appropriate framework based on three signals:
Relationship type — partner conflicts lean toward Gottman and attachment theory. Workplace conflicts map better to NVC and interest-based negotiation. Family dynamics often need a blend.
Conflict pattern — is this about communication style (Gottman), unmet needs (NVC), cognitive distortions (CBT), or deep-rooted attachment patterns? The pattern determines the lens.
Escalation level — high-escalation conflicts get de-escalation-first approaches. Low-stakes disagreements can jump straight to resolution frameworks.
This isn't a simple if-else tree. The classifier weighs multiple signals and sometimes combines frameworks. A partner conflict about household responsibilities might get Gottman for the communication breakdown analysis and NVC for the resolution path.
Prompt architecture per framework
Each framework has its own prompt template, but they share a common output structure. This was a deliberate design choice — users get consistent verdicts regardless of which framework runs under the hood.
The template structure looks like this:
Context injection — the user's conflict description, parsed into structured elements (who said what, what happened, what's the recurring pattern).
Framework application — the specific psychological model's lens applied to those elements. For Gottman, this means identifying which of the Four Horsemen are present. For NVC, it means extracting the unspoken needs behind each person's position.
Verdict generation — a clear assessment of what's happening, why it's happening, and what each person can do about it. Not "you're both right" — an actual analysis with reasoning.
Next steps — three to five specific, actionable things each person can try. Not vague advice like "communicate more." Concrete moves like "next time you notice yourself withdrawing from the conversation, say 'I need five minutes to collect my thoughts' instead of going silent."
Where AI fails — and how to catch it
The biggest quality issue isn't hallucination in the traditional sense. It's false confidence in framework application. The model will confidently apply Gottman's contempt pattern to something that's actually just frustration, because the surface-level language looks similar.
I built validation checks at three points:
Pre-generation — does the conflict description contain enough information to apply the selected framework? If someone writes "we argued about dinner," that's not enough signal. The system asks follow-up questions instead of guessing.
Post-generation — does the verdict actually use the framework correctly? I check for common misapplications: confusing criticism with contempt (Gottman), confusing observations with evaluations (NVC), or applying attachment labels without behavioral evidence.
User feedback loop — after each verdict, users can flag whether the analysis felt accurate. This doesn't retrain the model, but it feeds into my understanding of where each framework template needs refinement.
The hardest lesson: AI is surprisingly good at describing psychological concepts but mediocre at applying them to specific situations. The gap between "can explain Gottman" and "can correctly identify contempt in this specific conversation" is wider than you'd expect.
What I'd do differently
If I started over, I'd begin with fewer frameworks. I launched with five, and the interaction effects between them created edge cases I'm still sorting out. Start with one, nail the prompt template, validate the output quality, then add the next.
I'd also invest earlier in the classification layer. Getting the right framework matters more than perfecting any single framework's prompt. A mediocre NVC analysis of a situation that actually needs NVC beats a perfect Gottman analysis applied to the wrong conflict type.
This is part of what I'm building at Inithouse — a portfolio of AI-powered tools where each product tackles a specific problem domain. VerdictBuddy handles conflict resolution, Be Recommended tracks AI visibility for brands, and Here We Ask explores community-driven Q&A.
The psychology-to-prompt pipeline is the kind of problem that looks simple until you're three frameworks deep and your classifier is arguing with itself. But when it works — when someone gets a verdict that genuinely helps them understand their conflict — the architecture disappears and the output speaks for itself.
Top comments (0)