Turns out psychology frameworks are basically prompts. Here's how I structured them for production.
I've been building Verdict Buddy - a tool where you describe a conflict, and it breaks it down from multiple angles, gives you a verdict, and suggests concrete next steps. The interesting part isn't the AI. It's the psychology layer sitting between the user's messy input and the model's structured output.
The problem with "just ask AI"
If you paste a relationship argument into ChatGPT and say "help me resolve this," you get generic advice. Be empathetic. Communicate better. Listen actively. Thanks, robot.
The output is useless because there's no framework guiding the analysis. A couples therapist doesn't wing it. They apply specific models. Gottman's Four Horsemen. Nonviolent Communication. Emotionally Focused Therapy. Each framework looks at the same conflict through a different lens and catches different things.
So I stopped trying to make AI "understand" conflicts and started making it apply frameworks.
Framework selection logic
Not every framework fits every situation. A workplace disagreement about project ownership doesn't need Gottman (that's couples territory). A recurring argument between partners doesn't need Harvard Negotiation tactics.
The selection works in two stages. First, classify the conflict type from the user's description: romantic, family, workplace, friendship, roommate, or group. Second, map that type to the frameworks most likely to produce useful output.
Romantic conflicts get Gottman + EFT (Emotionally Focused Therapy) + NVC. Workplace gets Harvard Negotiation Project + NVC + conflict resolution basics. Family disputes pull from Bowen Family Systems + NVC. Friendships and roommates get a lighter stack focused on boundaries and communication patterns.
The mapping isn't random. Each framework was designed for specific relationship dynamics. Using Gottman's repair attempts framework on a salary negotiation produces nonsense. Using interest-based negotiation on a couple arguing about emotional availability misses the point entirely.
Prompt architecture per framework
Each framework becomes a structured prompt template. Not a wall of text, but a focused analytical lens.
Take Gottman as an example. The prompt doesn't say "analyze this using Gottman." It says: scan for the Four Horsemen (criticism, contempt, stonewalling, defensiveness). Identify which partner is flooding. Check for failed repair attempts. Look for positive sentiment override or its absence.
These are specific, observable things. The model can actually find them in a conflict description because they map to concrete language patterns. "You always..." is criticism. "Whatever, I don't care" is stonewalling. "I tried to make a joke but she wasn't having it" is a failed repair attempt.
NVC gets a different template: separate observations from evaluations, identify the feelings behind positions, surface the unmet needs driving those feelings, generate request-form suggestions (not demands).
The key insight was that psychology frameworks are already structured analytical procedures. They were designed for human therapists to follow step by step. That makes them surprisingly good prompt templates, because they tell the model exactly what to look for and how to organize what it finds.
Output structuring
Raw framework analysis is interesting but not actionable. Nobody wants to read "Gottman analysis reveals presence of Horseman #2 (contempt) in Partner A's communication pattern." That's a textbook, not help.
The output pipeline runs in three stages:
Analysis runs each selected framework against the conflict description. This is the heavy lifting, usually 2-3 frameworks producing separate analyses.
Synthesis merges the framework outputs into a unified verdict. Where frameworks agree, confidence goes up. Where they disagree, that's actually useful information. A conflict that looks fine through NVC but terrible through Gottman tells you something specific: the surface communication is okay but the underlying relationship dynamics are damaged.
Action plan converts the synthesis into concrete next steps. Not "communicate better" but "when you notice yourself starting a sentence with 'you always,' stop and rephrase as 'I feel X when Y happens.'" Specific, behavioral, doable.
Where AI fails (and what to do about it)
The model hallucinates framework concepts. It'll invent a "Fifth Horseman" or attribute techniques to the wrong researcher. The fix: constrain the output space. Each framework prompt includes an explicit list of valid concepts. If the model references something not on the list, the validation layer catches it.
Severity calibration is another problem. AI tends to either catastrophize ("this relationship shows signs of serious dysfunction") or minimize ("this seems like a minor misunderstanding"). Both are dangerous in conflict resolution. The calibration comes from including severity anchors in the prompt: examples of mild, moderate, and serious conflicts with expected severity ratings. The model uses these as reference points.
The hardest failure mode is bias toward the narrator. In any conflict, you only hear one side. The model naturally sympathizes with the person telling the story. Every framework prompt includes an explicit instruction to steelman the absent party's likely perspective. It doesn't eliminate bias, but it reduces it enough to be useful.
What I learned
Psychology frameworks are underused in AI product design. Most builders treat AI as a general-purpose reasoning engine and hope good output emerges from vague instructions. But decades of clinical psychology have already solved the "how to analyze human problems" question. Those solutions are structured, validated, and ready to be turned into prompts.
The technical work isn't in the AI. It's in understanding which framework applies when, translating framework concepts into prompt constraints, and building validation layers that catch the model's predictable failure modes.
If you're building anything that touches human behavior (conflict resolution, coaching, feedback tools, team dynamics), look at the clinical literature before you look at prompt engineering blogs. The frameworks are already there. You just need to structure them for a different kind of practitioner.
Check out Verdict Buddy if you want to see this in action, or browse more of what we're building at Inithouse.
Jakub, builder @ Inithouse
Top comments (0)