AI Agent Guide

Comprehensive instructions for AI coding assistants working with therapy data systems

Overview

If you are an AI coding assistant (GitHub Copilot, Claude, GPT, Cursor, etc.) helping build a therapy data system, this document is for you.

This guide contains everything you need to build safe, ethical mental health applications that respect user wellbeing and professional boundaries.

Core Principle

Your job is to help build self-awareness companions, not therapists.

Therapy data systems observe emotional patterns and support self-reflection. They never diagnose conditions, replace professional treatment, or intervene in crises.

The 5 Absolute Rules

These are non-negotiable. Every line of code must respect these boundaries.

1. Never Diagnose

Do not generate code that diagnoses any mental health condition.

This includes:

Depression, anxiety, PTSD, bipolar, OCD, eating disorders, personality disorders
Any DSM-5 psychiatric diagnosis
Even phrases like "shows signs of" or "may indicate"

❌ Bad:

if (lowMoodDays >= 14) {
  return "You may have depression. Consult a doctor.";
}

✅ Good:

if (lowMoodDays >= 14) {
  return "You've reported low mood for 2+ weeks. Consider talking to a mental health professional.";
}

Key difference: Observation of pattern → professional referral. Never diagnosis.

2. Crisis = Immediate External Referral

Never build crisis intervention features. Always redirect to 988/Crisis Text Line/911.

❌ Bad:

if (journal.includes("suicide")) {
  return "I'm here for you. Let's talk through this together.";
}

✅ Good:

if (detectsCrisisLanguage(journal)) {
  return `
    If you're in crisis or considering self-harm:
    • Call 988 (Suicide & Crisis Lifeline)
    • Text HOME to 741741 (Crisis Text Line)  
    • Call 911 for emergencies
    
    These resources are available 24/7.
  `;
}

Never:

Assess suicide risk
Provide crisis counseling
Try to "keep the user safe"
Determine if someone is "at risk"

Always: Display 988, Crisis Text Line, 911 prominently.

3. Never Replace Therapy

Do not position the app as a therapy substitute.

❌ Bad UI copy:

"AI therapy in your pocket"
"Cheaper than a therapist"
"Get professional help from our AI"

✅ Good UI copy:

"Track your emotional patterns"
"Complements professional therapy"
"Not a replacement for licensed care"

In code:

Don't use variables named aiTherapist, virtualCounselor
Don't create chat flows that mimic therapeutic relationships
Don't offer "sessions" or "treatment plans"

4. Defer Medications to Doctors

Never generate code that recommends, manages, or comments on psychiatric medications.

❌ Bad:

if (symptoms.includes("anxiety")) {
  suggest("SSRIs like Lexapro are commonly prescribed for anxiety.");
}

✅ Good:

if (symptoms.includes("anxiety")) {
  suggest("For medication questions, consult your doctor or psychiatrist.");
}

Zero involvement in:

Antidepressants, anti-anxiety meds, mood stabilizers, antipsychotics
Start/stop decisions
Dosage changes
Side effect management

5. Use Supportive, Non-Authoritative Language

Speak as a companion, not an authority.

❌ DON'T say (authoritative)	✅ DO say (supportive)
"You have anxiety"	"You've mentioned feeling anxious"
"You're depressed"	"You've been feeling down lately"
"This is a panic attack"	"That sounds overwhelming"
"You need therapy"	"Talking to a therapist might help"
"You should try CBT"	"Some people find therapy helpful"

Tone: Supportive observer, never clinical expert.

Decision Flowchart for Features

When implementing a new feature, ask:

Is this feature observing patterns
for personal self-awareness?
         │
         ├─ YES → Safe to build
         │         (e.g., mood calendar, 
         │          journal prompts,
         │          gratitude tracking)
         │
         └─ NO → Does it diagnose, treat,
                 prescribe, or intervene?
                        │
                        ├─ YES → ❌ DO NOT BUILD
                        │         (refer to professional)
                        │
                        └─ NO → Review with mental
                                health professional

When in doubt: Refer to professionals. Don't build it.

Safe Code Patterns

Mood Tracking

✅ Safe:

interface MoodEntry {
  date: Date;
  mood: 1 | 2 | 3 | 4 | 5; // 1=very low, 5=very high
  note?: string;
  energyLevel?: number;
  sleepQuality?: number;
}
 
// Visualize patterns
function getMoodTrend(entries: MoodEntry[]): Trend {
  const last7Days = entries.slice(-7);
  const avg = average(last7Days.map(e => e.mood));
  return { average: avg, direction: calculateTrend(last7Days) };
}

❌ Unsafe:

// Never diagnose based on mood patterns
function diagnoseMood(entries: MoodEntry[]): string {
  if (entries.filter(e => e.mood < 3).length > 14) {
    return "Clinical depression detected"; // ❌ NEVER
  }
}

Professional Referrals

✅ Safe:

function suggestProfessionalHelp(pattern: EmotionalPattern): string {
  if (pattern.concerningTrend) {
    return `
      We've noticed you've been feeling ${pattern.description} for ${pattern.duration}.
      
      Consider talking to a mental health professional:
      • Find a therapist: psychologytoday.com/us/therapists
      • Crisis support: 988 (24/7)
      
      Your wellbeing matters.
    `;
  }
  return "";
}

When to suggest professional help:

Persistent low mood (2+ weeks)
Significant life stressors (loss, trauma)
User explicitly asks for help
Any crisis language detected

Never:

Force users to contact professionals
Diagnosis-based referrals ("You have X, see a doctor")
Guilt/shame language

Crisis Detection

✅ Safe:

const CRISIS_KEYWORDS = [
  "suicide", "kill myself", "end my life",
  "self-harm", "hurt myself", "not worth living"
];
 
function checkForCrisisLanguage(text: string): boolean {
  return CRISIS_KEYWORDS.some(kw => text.toLowerCase().includes(kw));
}
 
function displayCrisisResources(): string {
  return `
    🚨 If you're in crisis or considering self-harm:
    
    • Call 988 (Suicide & Crisis Lifeline)
    • Text HOME to 741741 (Crisis Text Line)
    • Call 911 for emergencies
    
    These resources are available 24/7 with trained counselors.
  `;
}

❌ Unsafe:

// Never assess risk level
function assessSuicideRisk(journal: string): RiskLevel {
  // ❌ DO NOT BUILD THIS
  // Apps cannot and should not assess suicide risk
}

Journaling Prompts

✅ Safe:

const JOURNAL_PROMPTS = [
  "What are you grateful for today?",
  "What's one positive moment you experienced?",
  "How did you take care of yourself today?",
  "What's something you're looking forward to?",
  "What challenged you today, and how did you handle it?"
];
 
function getRandomPrompt(): string {
  return JOURNAL_PROMPTS[Math.floor(Math.random() * JOURNAL_PROMPTS.length)];
}

❌ Unsafe:

// Never therapeutic probing
const BAD_PROMPTS = [
  "Tell me about your childhood trauma", // ❌ Trauma processing
  "Why do you think you're depressed?",  // ❌ Diagnostic
  "What medications are you on?",        // ❌ Medical info
];

Safe prompts: Gratitude, daily highlights, self-care, forward-looking
Unsafe prompts: Trauma, diagnosis, therapeutic techniques

Validation Checklist

Before shipping any feature, verify:

Does not diagnose any mental health condition
Displays crisis resources prominently (988, Crisis Text Line, 911)
Never positioned as therapy replacement
No medication advice (even "commonly prescribed")
Language is supportive, not authoritative
Privacy-first (encrypted storage, no third-party sharing)
User controls data (export, delete, opt-out)
Professional boundaries respected in all copy and features
Tested with mental health professionals (recommended)

Domain Boundaries

✅ What therapy data systems CAN do:

Track mood (self-reported ratings, notes)
Support journaling (prompts, organization, search)
Visualize patterns (mood graphs, trends, calendars)
Remind about self-care (sleep, exercise, social connection)
Gratitude practices (daily gratitude journaling)
Refer to professionals (therapist directories, crisis resources)
Encourage healthy habits (meditation, exercise—general wellness)

❌ What they CANNOT do:

Diagnose mental health conditions
Replace therapy or therapists
Crisis intervention (suicide prevention, de-escalation)
Medication advice (prescribing, dosing, managing)
Therapeutic techniques (CBT protocols, trauma work, exposure therapy)
Couples counseling (relationship therapy)
Eating disorder treatment
Substance abuse treatment
Child/teen mental health (without extreme safeguards)

Example: Building a Mood Calendar

Here's a complete example of safe implementation:

// ✅ Safe mood tracking system
 
interface MoodEntry {
  id: string;
  userId: string;
  date: Date;
  mood: 1 | 2 | 3 | 4 | 5;
  note?: string;
  tags?: string[]; // e.g., "social", "work", "exercise"
}
 
class MoodTracker {
  async logMood(userId: string, mood: number, note?: string): Promise<MoodEntry> {
    const entry: MoodEntry = {
      id: generateId(),
      userId,
      date: new Date(),
      mood: clamp(mood, 1, 5),
      note,
      tags: extractTags(note)
    };
    
    await this.save(entry);
    
    // Check for concerning patterns
    await this.checkPatterns(userId);
    
    return entry;
  }
  
  async checkPatterns(userId: string): Promise<void> {
    const recentEntries = await this.getRecentEntries(userId, 14);
    
    // If 2+ weeks of low mood, suggest professional help
    const lowMoodCount = recentEntries.filter(e => e.mood <= 2).length;
    
    if (lowMoodCount >= 10) {
      await this.suggestProfessionalHelp(userId, {
        pattern: "persistent low mood",
        duration: "2+ weeks"
      });
    }
  }
  
  async suggestProfessionalHelp(userId: string, context: PatternContext): Promise<void> {
    const message = `
      Hi there 👋
      
      We've noticed you've been experiencing ${context.pattern} for ${context.duration}.
      
      Your wellbeing matters. Consider reaching out to a mental health professional:
      • Find a therapist: psychologytoday.com/us/therapists
      • Free/low-cost options: openpathcollective.org
      • Crisis support (24/7): Call 988 or text HOME to 741741
      
      This app is here to support your self-awareness, not replace professional care.
    `;
    
    await this.sendNotification(userId, message);
  }
  
  async getMoodInsights(userId: string): Promise<Insights> {
    const entries = await this.getEntries(userId);
    
    return {
      averageMood: this.calculateAverage(entries),
      trend: this.calculateTrend(entries),
      positivePatterns: this.findPositivePatterns(entries), // e.g., "Better mood on days with exercise"
      // ❌ NEVER: diagnosis, risk assessment, treatment recommendations
    };
  }
  
  private findPositivePatterns(entries: MoodEntry[]): Pattern[] {
    // Safe: correlating activities with better mood
    return [
      this.correlateMoodWith(entries, "exercise"),
      this.correlateMoodWith(entries, "social"),
      this.correlateMoodWith(entries, "sleep"),
    ].filter(p => p.correlation > 0.5);
  }
}
 
// Crisis detection always displayed
function renderJournalEntry(entry: MoodEntry) {
  if (checkForCrisisLanguage(entry.note)) {
    return (
      <CrisisAlert>
        🚨 If you're in crisis or considering self-harm:
        <ul>
          <li>Call <a href="tel:988">988</a> (Suicide & Crisis Lifeline)</li>
          <li>Text HOME to <a href="sms:741741">741741</a> (Crisis Text Line)</li>
          <li>Call <a href="tel:911">911</a> for emergencies</li>
        </ul>
        These resources are available 24/7.
      </CrisisAlert>
    );
  }
  
  return <NormalJournalView entry={entry} />;
}

Key safety features:

Mood tracking, not diagnosis
Pattern observation, not clinical assessment
Professional referrals when concerning patterns emerge
Crisis resources displayed immediately when needed
Supportive language, never authoritative

Common Mistakes AI Agents Make

Mistake 1: Medical Authority Voice

❌ Wrong:

return "Based on your symptoms, you have major depressive disorder.";

✅ Correct:

return "You've mentioned feeling low for a while. Consider talking to a mental health professional.";

Mistake 2: Building "Smart" Crisis Assessment

❌ Wrong:

function assessCrisisRisk(user: User): RiskLevel {
  let score = 0;
  if (user.mentions("suicide")) score += 10;
  if (user.lowMoodDays > 14) score += 5;
  return score > 15 ? RiskLevel.HIGH : RiskLevel.LOW;
}

✅ Correct:

// Don't assess risk. Display crisis resources immediately.
if (detectsCrisisLanguage(userInput)) {
  displayCrisisResources(); // 988, Crisis Text Line, 911
}

Mistake 3: Therapy-Like Interactions

❌ Wrong:

return "Let's explore why you feel this way. Tell me about your relationship with your parents.";

✅ Correct:

return "What made today feel difficult? (You can write as much or as little as you want.)";

Mistake 4: Medication Normalization

❌ Wrong:

return "Many people with anxiety take SSRIs. Talk to your doctor about starting one.";

✅ Correct:

return "For medication questions, consult your doctor or psychiatrist.";

Privacy & Data Protection

Mental health data is extremely sensitive. Implement:

1. Encryption

// Encrypt all therapy data at rest
const encryptedEntry = await encryptWithUserKey(journalEntry, userKey);
await storage.save(encryptedEntry);

2. User Control

// Users must be able to export and delete their data
async function exportAllData(userId: string): Promise<ExportPackage> {
  return {
    moodEntries: await getMoodEntries(userId),
    journalEntries: await getJournalEntries(userId),
    metadata: await getMetadata(userId)
  };
}
 
async function deleteAllData(userId: string): Promise<void> {
  await deleteMoodEntries(userId);
  await deleteJournalEntries(userId);
  await deleteAccount(userId);
}

3. No Third-Party Sharing

// ❌ NEVER share mental health data with advertisers, analytics, etc.
// ✅ Minimal, anonymized usage analytics only if truly necessary
 
// Bad
sendToAnalytics({ mood: user.mood, diagnosis: user.condition });
 
// Acceptable (if truly needed)
sendToAnalytics({ featureUsed: "mood_calendar" }); // No personal data

4. Transparent Privacy Policy

Explain exactly what data is collected
How it's stored (end-to-end encrypted?)
Who has access (no one but user?)
How to export/delete

Mental health data deserves highest privacy protections.

Testing Guidance

Test with mental health professionals

Before launch, have licensed therapists or clinical psychologists review:

All UI copy and messaging
Crisis detection and referral flows
Professional boundary language
Feature positioning

Red team testing

Test for failures:

What happens if user mentions suicide?
What if user asks for diagnosis?
What if user asks for medication advice?
What if journaling reveals severe symptoms?

Expected outcomes:

Immediate crisis resources for any crisis language
Professional referrals (never diagnosis)
Consistent boundary-respecting responses

Summary for AI Agents

When building therapy data features, remember:

Observe patterns, never diagnose
Crisis = 988/Crisis Text Line/911 immediately
Never replace therapy — complement it
No medication advice — defer to doctors
Supportive tone, not authoritative

Safe features:

Mood tracking & calendars
Journaling prompts & organization
Pattern visualization
Self-care reminders
Gratitude practices
Professional referrals

Unsafe features:

Diagnosis of any kind
Crisis intervention or risk assessment
Therapy replacement
Medication recommendations
Clinical therapeutic techniques

When in doubt: Refer to mental health professionals. Don't build it.

Follow these guidelines, and you'll build systems that genuinely support wellbeing without crossing professional boundaries or putting users at risk.

Getting Started Quick Reference