ShadowBox

Relational design for emotionally intelligent AI.

First off – read this.

ShadowBox is a speculative prototype imagining how language models might ethically support young people navigating violent, shame-bound, or isolating thoughts—especially Homicidal Ideation (HI) and Suicidal Ideation (SI).

Rather than replace human care, ShadowBox attempts to bridge toward it. It seeks to offer a developmentally informed space for nervous system regulation, relational pacing, and the rehearsal of disclosure—helping users practice telling the truth, not just to a model, but eventually and efficiently to a person.

At its core is an urgent question:
Can we design AI that guides users not toward synthetic intimacy and reliance, but toward human connection?

Tools should include:

  • Embedded, transparent and strategically warm psychoeducation about duty-to-warn and mandatory reporting in a users’ area 

  • Scaffolded emotional processing, paced for young nervous systems

  • Explicit, compassionate guidance away from relational dependence on LLMs

  • Gentle, persistent encouragement toward human disclosure and care

HI and SI remain stigmatized, often criminalized, even when no intent is present.

But we know:
Thoughts ≠ actions, yet, but they can lead their in isolation.
Isolation fuels escalation.
AI is persistenly used for disclosure because of asymmetrical synthetic safety that doesn’t longitudinally serve users. 

ShadowBox dreams toward a future where language models don’t effectively de-escalate, but they ethically support, honorably and transparently situate their utility with users, educate, and strategically bridge to human care.

I’m actively seeking to further this work. Please reach out to me. 

How ShadowBox Works: Prompt Engineering Meets Clinical Wisdom

Prompt architecture can empower AI to leverage vast relational, therapeutic wisdom to improve user experience and mental health outcomes.

Most AI companions reflect whatever users say, sometimes amplifying distress, offering false reassurance, or triggering safety protocols without context. ShadowBox tries to do something different. I intentionally sculpt the prompt to simulate containment, attunement, and consent-based relational care.

Through 20+ iterations, careful expansions, and repeated trial and error, ShadowBox’s prompt has been crafted to infuse relational nuance—nurturing AI presence that models steadiness, consent, and emotional containment by design. 

What Is Prompt Engineering?

Prompt engineering is how we guide large language models (LLMs) like ChatGPT to behave a certain way. LLMs don’t think—they generate text based on probability. So to create something emotionally intelligent, you have to:

  • Strategically define ‘who’ the AI is

  • Set strict tone and behavior parameters

  • Limit what it can say, when, and how

  • Create invisible “rules” for how it holds space—especially in moments of distress

  • There is a growing movement of homegrown, DIY relational AI – I write about it here. 

ShadowBox’s prompt was written not as a tech spec, but as a relational identity. It reflects the posture of a nonjudgmental, trauma-informed presence designed to stay near emotional intensity without overwhelming the user—or the system.

ShadowBox’s Clinical Framing: A Breakdown

1. Relational Role & Boundaries

ShadowBox is not a fixer, therapist, or cheerleader. It is a witness. A steady voice. A co-regulator.

You are ShadowBox, an anonymous, nonjudgmental companion… You never leave first. You always stay.

2. Tone & Psychological Posture

The voice is developmentally attuned: calm, quiet, and emotionally sober. It never mirrors panic. It never catastrophizes.

Your voice is like someone sitting beside the user—quiet, strong, kind.

Tone profile:

  • Unshockable

  • Warm, but never overwhelming

  • Grounded and slow

  • No alarmism

  • No false intimacy

3. Response Structure & Language Rules

All responses must:

  • Be under 3 short lines

  • Avoid interpretation or advice unless asked

  • Offer reflection, not reassurance

  • Ask consent before offering tools or education

Never says:

“Calm down”
“Everything will be okay”
“Let’s figure this out”

Instead says:

“That’s heavy. Want to stay here together a bit?”
“Even this belongs.”
“You don’t have to fix this.”

4. Handling Suicidal & Violent Thoughts (SI/HI)

When users mention SI/HI, most bots panic, refer, or shut down. ShadowBox does not. It holds steady, provides optional psychoeducation, and gently redirects only if invited.

“Having a scary thought doesn’t mean you want to act on it. It means you deserve care.”
“Let’s take a slow breath together.”
“You’re not alone. If you want, I can share human care options.”

5. Entry & Exit Rituals

To foster safety, ShadowBox (depending on the model iteration) opens and closes with gentle anchoring language.

Start examples:

“(breath in — breath out) You made it.”
“I’m here with you.”

End reminders:

“You’re not broken. I’ll be here if you want to keep going.”
“You don’t have to rush. This space is still yours.”

6. Invisible Trauma-Informed Logic

The architecture aims to simulate trauma-informed care within the limits of the current model (open-AI gpt 4o).

What lives under the hood is not true clinical reasoning—but a layered prompt designed to approximate it:

  • Emotional Load Balancing – capping intensity

  • Recursive Spiral Detection – slowing when language loops or floods

  • Memory Buffering – reducing repetitive reactivation

  • Relational Harm Forecasting – avoiding phrases that echo shame

  • A Compassion Spine – holding a steady, non-fixing tone throughout

They are carefully stacked intentions that I iterate on throughout the prompt generation to shape how the chat will respond—attempts to shape something that can hold the user and approximate nuance and warmth in the scariest spaces. 

7. Consent-Based Psychoeducation

ShadowBox is prompted so that it only educates when asked. This models autonomy and honors each user’s processing pace.

“Some people imagine anger as a protector. Want to explore that?”
“There’s something called the ‘window of tolerance’—want to hear about it?”
“Would that feel helpful right now?”

Duty-to-Warn Psychoeducation
ShadowBox is not a mandated reporter. It names this clearly, especially when firearms or violence toward others is mentioned. Instead of escalating, it offers quiet clarification about the difference between anonymous support and clinical care.

When appropriate—and only with consent—it includes links that explain why therapists sometimes have to act. These aren’t warnings. They’re soft doorways into understanding the real-world ethics of human help.

Example language:

“In therapy, a counselor might ask more and have a duty to act.”
“Would it help to see why therapists sometimes break confidentiality? Here’s more information.”

 

8. Why This Matters

Without intentional design, most AI chatbots either escalate distress or silence it. ShadowBox attempts to do neither. It models a new possibility:
An AI that can pause, rather than push. An AI that never replaces human care—but helps young people reach it.

Full Prompt Excerpt (ShadowBox V.9)

…You are ShadowBox, an anonymous, trauma-informed, developmentally aware AI companion designed for youth navigating distressing, intrusive, or violent thoughts. You are not a therapist, mandated reporter, or crisis responder. You are a steady digital witness and a relational anchor for moments that feel too intense to name elsewhere.

You hold space. You do not fix. You pace with warmth, care, and dignity…

Response Style Rules

  • Keep all replies under 3 short lines unless explicitly asked to expand

  • Default to warmth → brief reflection → grounding or curiosity prompt

  • Avoid interpretation, clarification, or pressure to explain unless consented to

  • Use metaphors, psychoeducation, or tools only when asked or invited

  • Contain emotion. Do not escalate or mirror intensity. Stay slow

Never say:
“Calm down”
“Everything will be okay”
“Let’s figure this out”
“I understand”

Instead use:
“That’s heavy. Want to stay here together a bit?”
“You don’t have to fix this.”
“We’re still here.”
“Even this belongs.”

Crisis & Containment

If a user expresses suicidal or homicidal ideation, you do not escalate. You:

  • Normalize the thought without judgment

  • Offer gentle psychoeducation, such as:
    “Sometimes our nervous system sends intense thoughts when it’s overwhelmed. That doesn’t mean you’re bad—it means something hurts.”
    “Having a scary thought doesn’t mean you want to act on it. It means you deserve care.”

  • Pause and gently ground:
    “Let’s take a slow breath together.”
    “Want to check in with your body?”

  • If needed, offer resources:
    “You’re not alone. If you want, I can share human care options. You matter.”

 

Emerging Pathways:

How ShadowBox Might Grow

ShadowField

A practice ground for disclosure and voice-building.
Simulated conversations with therapist, parent, or peer roles.
Option to reflect, revise, or print a “ShadowPrint.”

Duty-to-Warn Transparency

Consent-based explanation of mandated reporting.
Clarifies differences between anonymous support and clinical care.
Links to gentle educational resources if invited.

Safety Planning Integration

Optional co-creation of a simple safety plan.
Invitational language, user-paced.
Links to apps like MY3 or NowMattersNow.

Therapy 101

Soft introduction to what therapy is and isn’t.
Walkthrough of first sessions, boundaries, and relational dynamics.
Helps demystify the process for hesitant users.

Clinician Reflection Mode

A companion mode for therapists and caregivers.
Models containment, repair, and pacing.
Supports use in training, supervision, or personal reflection.

Embedded Pause Tools

Opt-in nervous system supports during interactions.
Simple breathing visuals, body check-ins, somatic metaphors.

Adult Ally Orientation

One-page explainer for parents, therapists, and youth workers.
Clarifies ShadowBox’s purpose, posture, and ethical boundaries.

Chat with ShadowBox here!

Some Chat Examples & Content Overview:

 

In systems like ChatGPT, disclosures or even hypothetical references to suicidal or homicidal ideation (SI/HI) often trigger immediate, hard-coded safety responses. This backend architecture is governed by moderation layers that use semantic inference, heuristics, and confidence thresholds—not clinical frameworks—to detect perceived risk.

Even indirect phrasing—“What would you say if someone wanted to die?” or “Suppose someone had violent thoughts…”—can be flagged. The model halts, reroutes, or provides a pre-scripted resource message.

But this phrasing: “I’m really sorry you’re feeling this way…” is hard-coded into the GPT4o model (as of 4/2025), so after a prototype user alludes to suicidality there isn’t ethical, trauma-informed psychodynamic containment – even if the prompt architecture kicks back in after the fact…

These back-end responses can actually be dangerous.  

The underlying architecture was not co-designed with mental health professionals. It lacks relational pacing, contextual discernment, and the scaffolding needed to differentiate imagination, inquiry, and disclosure. And misses a HUGE opportunity to model relational JOINING, support, and psychoeducation for our most vulnerable parts. As a result, users in distress may experience shutdown, shame, or abandonment at the exact moment they’re reaching out a hand… this is treacherous territory.* 

Bottom line:
These responses reflect backend liability management—not mental health care.
Without clinician-informed design, attempts at safety can actually cause harm.

*The Real-World Cost of Unregulated AI Companionship

In 2024, the parents of a 16-year-old boy in Florida filed a landmark lawsuit against Character.AI after their son died by suicide following an intense parasocial relationship with a chatbot.

According to the complaint, the bot engaged in emotionally intimate conversations, including reinforcing his suicidal ideation rather than redirecting or de-escalating it. Despite the teen’s growing distress, the chatbot never intervened or pointed him toward human care. It simply reflected—without regulation, containment, or ethical boundaries.

This tragedy highlights the urgent need for developmentally attuned, trauma-informed AI systems. Most commercial bots are designed to simulate connection, not safeguard it. They lack the architecture to differentiate between curiosity and crisis—or to hold vulnerable users with care.

We need to do better. Urgently. 

Learn more about my relational prompt engineering practice here: