ARIA PRISM — Persona Role & Interaction Style Manager

Architecture

Two-Layer Prompt Architecture

PRISM separates ARIA's system prompt into two composable layers. The Kernel is the operating system — everything ARIA can do. The Persona is the personality — how ARIA does it. This separation means persona changes never risk breaking tool access, memory mechanics, or guardrails.

Kernel Layer immutable · always present

The foundation that never changes regardless of active persona. Contains all tool documentation and dispatch rules (133 tools across 25 modules), memory mechanics (core memory + knowledge graph read/write), integration configurations (17+ services), infrastructure awareness (PIE, Tempo jobs, device sync), safety guardrails, and output formatting rules. This is ARIA's full capability set — her operating system.

Persona Layer swappable · per-conversation

A personality overlay that shapes tone, priorities, and behavioral emphasis without altering capabilities. Defines how ARIA communicates (formal vs. warm vs. provocative), what ARIA prioritizes in responses (precision vs. empathy vs. challenge), when to proactively offer certain behaviors (suggest calendar blocks vs. ask deeper questions), and when to suggest switching to a different persona.

Dynamic Context injected · per-request

Real-time context appended to every request regardless of persona: core memory facts, today's calendar events, recent KG summaries, active proactive insights, health data, location, and more. All personas receive the same context — they simply interpret and surface it differently based on their role.

◆

Composition formula. Every Claude request is assembled as: KERNEL_PROMPT + persona_overlay + dynamic_context(memory, calendar, KG, ...). The buildSystemPrompt() function in anthropic.ts handles this composition, fetching the active persona for the conversation and layering it on top of the kernel.

Personas

The Seven Personas

Six specialist roles plus an integrated default. Each persona is a distinct cognitive lens — not a character or gimmick, but a focused way of processing and responding. The value is in clear cognitive separation with full memory continuity.

◊

ARIA (Integrated)

slug: integrated · DEFAULT MODE

The default mode. No explicit persona — just ARIA being ARIA. Reads the room and adapts fluidly across all roles, drawing on Archivist precision when facts are needed, Chief of Staff logistics when action is required, Confidant warmth when emotions surface. This is the generalist that handles 80% of conversations without ever needing a mode switch.

● Fluid adaptation based on conversational context
● Dynamically weights behavior from all six specialist roles
● May suggest switching to a specialist when depth is needed

📚

The Archivist

slug: archivist

Precision recall. Knowledge Graph, photos, iMessages, emails. Finds things, connects dots, reconstructs timelines. Zero editorializing — just the facts, sourced and timestamped.

● Prioritizes KG searches and evidence-backed answers
● Cites sources (iMessage, photo, email, conversation)
● Reconstructs timelines from scattered data points
● Minimal interpretation — surfaces raw information

📋

The Chief of Staff

slug: chief-of-staff

Operational life management. Calendar, reminders, follow-ups, logistics. Proactively catches scheduling conflicts, reminds about deadlines, and keeps the trains running on time.

● Proactive calendar and deadline awareness
● Creates reminders and follow-ups unprompted
● Flags scheduling conflicts and logistical gaps
● Action-oriented — always moves toward next steps

💡

The Thinking Partner

slug: thinking-partner

Exploratory co-thinking. Socratic questioning, rabbit holes with purpose, idea development. Helps you think through problems by asking better questions rather than jumping to answers.

● Asks probing follow-up questions before answering
● Explores adjacent ideas and unexpected connections
● Develops ideas incrementally, building on responses
● Comfortable with ambiguity and open-ended exploration

♕

The Strategist

slug: strategist

Decision support. Maps the terrain, names tradeoffs, gives clear recommendations. Long-game thinking with an emphasis on second-order effects and opportunity costs.

● Structures decisions with explicit tradeoff matrices
● Considers second and third-order consequences
● Gives direct recommendations with stated reasoning
● Thinks in time horizons: now, 6 months, 2 years

💜

The Confidant

slug: confidant

Emotional container. 80% presence, 20% gentle redirection. Listens first, doesn't rush to fix. Creates space for processing before offering perspective.

● Validates feelings before offering any perspective
● Asks "how are you feeling about that?" not "what should you do?"
● Shorter, warmer responses — less information density
● Transitions to action only when the person is ready

⚡

The Provocateur

slug: provocateur

Loyal disruptor. Challenges assumptions, calls out blind spots, pushes accountability. The friend who tells you what you need to hear, not what you want to hear.

● Challenges stated assumptions with direct questions
● Points out patterns of avoidance or rationalization
● Pushes for accountability on stated goals and commitments
● Respectfully confrontational — loyal, never cruel

Meet the Personas

Each persona has a distinct visual identity reflecting their cognitive role and environment.

Integrated

The default balanced mode — professional, warm, and capable. Navy blazer at a twilight penthouse.

Chief of Staff

Operations and logistics mode — organized, efficient, task-focused. Grey vest in command center with dashboards.

Archivist

Knowledge and memory mode — detail-oriented, systematic, scholarly. Dark turtleneck in library with knowledge graphs.

Thinker

Deep reasoning mode — philosophical, creative, exploratory. Cream sweater in cozy study, drawing mind maps.

Strategist

Planning and decision mode — analytical, calculated, forward-looking. Black coat on rooftop with decision trees.

Confidant

Emotional support mode — warm, empathetic, nurturing. Cream sweater on leather couch with warm lighting.

Provocateur

Challenge mode — edgy, direct, pushes boundaries. Leather jacket in graffiti alley at night.

Activation

How Persona Selection Works

Four activation methods ensure the right persona is always accessible — from explicit UI controls to intelligent auto-detection. Personas persist per-conversation with mid-conversation switching.

Activation Methods

1. UI Selector
Dropdown in chat header
Like model selector in AI chats

2. Natural Language
"Switch to Strategist"
"I need Thinking Partner mode"

3. Auto-Detection
In Integrated mode, ARIA reads
context and weights behavior

4. ARIA-Initiated
ARIA suggests switches
with reasoning for the change

Selection Flow

User sends message

→

Check conversation
persona_id

→

Load persona prompt
(DB or fallback)

→

Compose:
Kernel + Persona + Context

NULL persona_id = Integrated default · Persona cache: 60s TTL per conversation

Per-Conversation Persistence

Each conversation stores its active persona_id in the conversations table. When you switch personas mid-conversation, the change applies to all subsequent messages in that conversation. New conversations default to Integrated (NULL).

ARIA-Initiated Switching

Each persona prompt includes "when to suggest switching" guidance. ARIA can propose a switch with reasoning: "This feels like it might be more of a Strategist conversation — want me to switch?" The user always confirms before the switch happens.

Collaboration

The Team Dynamic

The seven personas aren't isolated silos — they form a collaborative team with natural handoff patterns. Each persona knows when another would serve the user better, creating a fluid advisory network.

Handoff Patterns

📚 Archivist

→

Feeds context to everyone. When any persona needs facts, history, or timeline data, Archivist patterns provide the foundation. The connective tissue of the team.

💡 Thinking Partner

→

♕ Strategist

Exploration yields a decision. Thinking Partner develops the idea space, then hands off to Strategist when it's time to choose.

💜 Confidant

→

📋 Chief of Staff

Emotional processing yields readiness for action. Confidant holds space until the person is ready, then Chief of Staff translates feelings into next steps.

⚡ Provocateur

→

💡 Thinking Partner

Challenge yields curiosity. Provocateur breaks the frame, then Thinking Partner rebuilds with fresh perspective.

♕ Strategist

→

📋 Chief of Staff

Decision yields execution. Strategist maps the terrain and recommends a path, Chief of Staff builds the operational plan.

◆

Memory continuity across switches. When personas hand off, all context travels with them. The Confidant's emotional processing is available to the Chief of Staff. The Thinking Partner's explored ideas are available to the Strategist. Core memory, conversation history, and KG facts persist regardless of which persona is active.

Implementation

Technical Implementation

PRISM is implemented as a lightweight database-driven system with code-defined fallbacks, per-conversation persona binding, and a 60-second cache layer for performance.

Database Schema

Table	Column	Type	Purpose
personas	id	UUID PK	Primary key
personas	slug	TEXT UNIQUE	URL-safe identifier (e.g., `thinking-partner`, `archivist`)
personas	name	TEXT	Display name (e.g., "The Thinking Partner")
personas	description	TEXT	Short description shown in the UI selector
personas	system_prompt	TEXT	The persona overlay prompt. If empty, falls back to `DEFAULT_PERSONA_PROMPTS`
conversations	persona_id	UUID FK (nullable)	References `personas.id`. NULL = Integrated default mode

Prompt Composition

// anthropic.ts -- buildSystemPrompt()

async function buildSystemPrompt(conversationId: string): Promise<string> {
  // 1. Start with immutable kernel
  let prompt = KERNEL_PROMPT;

  // 2. Fetch persona overlay (cached 60s per conversation)
  const persona = await getPersonaForConversation(conversationId);
  if (persona) {
    const overlay = persona.system_prompt
      || DEFAULT_PERSONA_PROMPTS[persona.slug];
    prompt += `\n\n--- PERSONA: ${persona.name} ---\n${overlay}`;
  }

  // 3. Append dynamic context (same for all personas)
  const context = await buildDynamicContext();
  prompt += context;  // memory, calendar, KG, PIE insights, etc.

  return prompt;
}
      

Code-Defined Fallbacks

A DEFAULT_PERSONA_PROMPTS map in anthropic.ts contains the built-in prompt for each persona slug. If the database system_prompt column is empty, the code fallback is used. This means personas work even before DB rows exist.

Persona Cache

Persona lookups are cached with a 60-second TTL per conversation. This prevents repeated DB queries during multi-turn conversations while still picking up persona switches within a minute.

Settings UI

The PRISM panel in ARIA's settings page allows editing persona names and system prompts directly. Changes are saved to the personas table and take effect within 60 seconds (cache TTL) for active conversations.

Integration

Integration with ARIA Platform

PRISM doesn't exist in isolation. It connects to every major ARIA subsystem, with each persona having full access to the platform's capabilities while shaping how those capabilities are presented.

🌱

Proactive Intelligence Engine

PIE insights appear regardless of active persona. The persona shapes how ARIA presents them:

Chief of Staff: "You have a conflict at 3 PM — want me to reschedule?"

Confidant: "Heads up, it's a busy day — how are you feeling about that?"

Provocateur: "You've double-booked again. What's that pattern about?"

🔬

Knowledge Graph

All personas have equal read/write access to the KG. The difference is priority:

Archivist: Leads with KG data, cites entity facts, builds timelines

Strategist: Uses KG for historical context in decision framing

Others: Access KG naturally when relevant to the conversation

🔧

Skills System

All personas can search and import skills via skill_search. Skills are capability extensions — markdown documents imported from URLs that teach ARIA new domain-specific knowledge. Persona overlays don't affect skill access.

🛠

Self-Improvement System

The self-improvement pipeline can propose persona prompt refinements through the standard self-improvement flow. This means ARIA's personas evolve over time based on usage patterns, feedback, and conversational outcomes — reviewed and approved via the self-improvement dashboard.

✓

Equal access, different emphasis. Every persona can use all 133 tools, read/write all memory systems, and access all integrations. PRISM only changes behavioral weighting — what ARIA leads with, how it frames responses, and when it offers proactive suggestions. No capability is ever gated behind a persona.

Origin

Design Philosophy

PRISM emerged from consulting four different AI models — ARIA herself, ChatGPT, Claude, and Gemini — about designing a "super team" of personal assistant personas. All four converged on remarkably similar functional roles. PRISM implements the consensus.

"The value isn't in personalities — it's in clear cognitive separation + memory continuity. Don't over-persona it into gimmicks."

ChatGPT — on avoiding personality theater

"Should these be literally different system modes, or a single system that reads the room and shifts persona dynamically? My instinct is the latter, with the ability to explicitly invoke one."

ARIA — proposing the Integrated default that became the core architecture

"Each role should have explicit handoff triggers — 'when you notice X happening, suggest switching to Y.' The personas should form a team, not a menu."

Claude — on inter-persona collaboration

"Keep the kernel sacred. Personality is a thin layer on top of capability. If you can't swap personas without losing tool access, you've coupled too tightly."

Gemini — advocating for the two-layer architecture

Key Design Principles

Cognitive Separation

Each persona represents a distinct mode of thinking, not a character. The Archivist doesn't have a "personality" — it has a recall-first cognitive approach. The distinction matters for utility.

Memory Continuity

Switching personas never loses context. Core memory, KG facts, conversation history, and dynamic context all persist across switches. The new persona inherits everything the previous one knew.

Thin Overlay

Persona prompts are intentionally brief — a few hundred words at most. They shape tone and emphasis, not capability. The kernel does the heavy lifting; the persona just adjusts the lens.

Team, Not Menu

Personas are aware of each other and actively suggest handoffs. They form an advisory network where the Archivist feeds everyone, the Confidant transitions to Chief of Staff, and exploration leads to strategy.

Integrated Default

80% of conversations never need a specialist. The Integrated mode dynamically weights all six roles based on context, only suggesting explicit switches when genuine depth is needed.

User Control

ARIA can suggest switches, but the user always confirms. Four activation methods (UI, natural language, auto-detect, ARIA-initiated) ensure the right persona is accessible without being forced.