Hugging Face's logo

Spaces:
jostlebot
/
PromptPrism.Prototype
Running

App Files Community
PromptPrism.Prototype / README.md
jostlebot's picture
jostlebot
Add short description, trigger rebuild
b2e56a1
preview code
|
raw
history blame contribute delete
6.04 kB

A newer version of the Gradio SDK is available: 6.9.0

Upgrade
metadata 
title: Prompt Prism Prototype
short_description: Dynamic system prompts for psychiatric AI
emoji: 🔮
colorFrom: purple
colorTo: blue
sdk: gradio
sdk_version: 4.44.1
python_version: '3.11'
app_file: app.py
pinned: false
license: mit

Prompt Prism Prototype

Dynamic System Prompt Architecture for Psychiatric AI

Same AI foundation, different refractions based on clinician + client + session.

The Problem

Psychiatric AI tools today give every clinician the same prompt.

  • A psychodynamic therapist gets CBT-flavored AI
  • A high-risk bipolar client gets the same flags as stable anxiety
  • Intake data sits unused in the EHR
  • Clinician expertise is flattened to one-size-fits-all

The Solution

Prompt Prism generates a unique system prompt for each clinician + client combination.

Same AI infrastructure. Different refraction for each clinical dyad.

How It Works

The Prism Metaphor

A prism takes one light source and refracts it differently based on angle.

Prompt Prism takes the ARI Framework (ethical base layer) and refracts it based on:

  • Clinician (orientation, style, philosophy)
  • Client (risk, diagnosis, treatment stage)
  • Tools (opt-in modules enabled for this client)
  • Session (what's happening now)

The Layers

Layer What It Contains Why It Matters
ARI Base Ethical guardrails, crisis protocols, forbidden language Always-on safety
Clinician Orientation, style, preferences, exclusions AI thinks like you do
Client Risk level, diagnosis, stage, custom flags Personalized care
Tools Opt-in modules clinician enables Right tools for right client
Session Last session, today's focus, current state Continuity

Opt-In Tools

Clinicians enable specific tools for specific clients:

Tool Purpose
Diagnosis Explorer Multi-pillar psychoeducation without pathologizing
Tend & Send NVC-based communication for couples/family
NVC How-To Nonviolent Communication skill-building
Distress Tolerance DBT skills for riding out intensity
Practice Conversations Roleplay difficult discussions
Grounding & Regulation Somatic/sensory grounding
Somatic Check-In Brief interoception practice
ShadowBox Static Library Crisis-adjacent psychoeducation (NO LLM - static content only)

About ShadowBox

ShadowBox demonstrates how to handle crisis-adjacent content safely:

  • Pre-written, clinically-reviewed content only
  • NO LLM generation for suicidal ideation, self-harm, or crisis topics
  • Confidentiality explained clearly
  • State-specific duty-to-warn information
  • Safety planning (Stanley-Brown model)
  • Crisis resources with context
  • Starter scripts for disclosure

"A resonant library for hard thoughts. Not a chatbot."

Supported Orientations

  • Psychodynamic
  • CBT
  • DBT
  • Trauma-Informed
  • IFS (Internal Family Systems)
  • Somatic/Body-Based
  • ACT (Acceptance and Commitment)
  • Humanistic/Person-Centered
  • Integrative

Built On: ARI Framework

Assistive Relational Intelligence - AI that scaffolds human connection, not simulates it.

Core Principles

  1. Scaffold human connection, not simulate it
  2. Bridge toward human care, not away from it
  3. Build capacity, not dependency
  4. Honor clinician expertise
  5. Refuse engagement-optimization

Safety Features

  • Forbidden language patterns (no synthetic intimacy)
  • Crisis protocols that defer to humans
  • Risk-calibrated alerting
  • Session boundaries and exit rituals
  • Always bridge back to psychiatrist

For Psychiatric Organizations

This prototype demonstrates how hundreds of psychiatrists could each get AI aligned to their practice:

  • Psychodynamic psychiatrist → psychodynamic AI responses
  • DBT psychiatrist → DBT-aligned skill coaching
  • Conservative risk philosophy → lower alert thresholds
  • Client in stabilization → no trauma processing suggested
  • Couples client → Tend & Send communication tools enabled
  • High-risk client → ShadowBox static library for safe psychoeducation

The result: AI that feels like an extension of the clinician's approach, not a generic chatbot.

What This Demonstrates

For each unique clinician-client relationship:

  1. Provider-controlled UX — Psychiatrist toggles which tools are available
  2. Modality alignment — AI speaks in the clinician's orientation
  3. Risk calibration — Alerts tuned to this client's specific profile
  4. Opt-in tooling — Right tools for right client at right time
  5. Bridging architecture — Every tool points back to human care
  6. Static content for crisis — ShadowBox shows how to handle high-risk topics safely

Innovation points:

  • Dynamic system prompt generation per dyad
  • Clinician as configurator, not just consumer
  • Human-in-the-loop at every layer
  • Ethical guardrails embedded, not bolted on
  • Between-session support that strengthens (not replaces) the psychiatric relationship

Usage

  1. Clinician Profile - Set your orientation, style, preferences
  2. Client Context - Configure diagnosis, risk level, treatment stage
  3. Tools & Modules - Enable opt-in tools for this client
  4. Session Context - Add continuity from last session
  5. Generate Prompt - See your compiled system prompt
  6. Test Prompt - Try it with sample client messages

Setup

Add your ANTHROPIC_API_KEY in Space settings to enable the Test Prompt feature.

Author

Jocelyn Skillman, LMHC

Clinical AI Designer | Creator of the ARI Framework

"Your 800 psychiatrists practice differently. Should they all get the same AI?"

Related Projects