The Council of Advisors: Building a Multi-Persona AI Layer Into a Personal Dashboard

I asked Gemini to assemble a council of advisors for me. Not a committee. A cabinet. Eight distinct personas, each with a name, a personality, and a specific lens through which they interpret the world. What came back was surprisingly useful, and it got me thinking about what it would take to actually build this into Life Hub.

The concept: instead of one generic AI assistant that does everything in the same monotone, you have a team. Alistair Vance is the cold strategist who only cares about optimal moves. “Birdie” Malone is the chaotic wildcard who finds the path your spreadsheets would never catch. Coach Miller watches your sleep score and forces you to drink water. Bramwell Finch’s entire job is to argue the opposite of whatever everyone else just agreed on.

This is not a toy idea. This is an architectural pattern.

The Council

Eight advisors, each tuned to a different dimension of decision-making:

The flaws matter. An advisory council where everyone agrees with you is just an echo chamber with better production values. Bram exists specifically to find the hole in Alistair’s plan. Miriam exists to ask whether Jax’s “efficient” solution is one you can actually live with.

Why Personas Beat a Single Assistant

I have been using Claude daily through Life Hub’s MCP server for months. It logs habits, checks infrastructure, queries sleep patterns, writes digest summaries. One assistant, one voice, one mode.

The limitation is not capability. It is perspective. When I ask Claude about a scheduling conflict, it gives me the same analytical tone it uses for debugging Kubernetes manifests. When I need a creative pivot on a stuck project, it gives me the same measured response it uses for writing governance documents.

Personas solve this by routing context through different system prompts. Same underlying model, radically different outputs. Alistair gets a system prompt emphasizing data-driven analysis, probability assessment, and zero sentimentality. Birdie gets one emphasizing lateral thinking, pattern-breaking, and creative risk. Coach Miller gets one focused on physiological data, recovery science, and behavioral nudges.

The model does not change. The lens does.

The Architecture: MCP as the Nervous System

Life Hub already runs an MCP server that exposes 14 tools across habits, mood, sleep, goals, journal, health metrics, infrastructure monitoring, reading stats, music activity, and GitHub contributions. Seven collectors pull data from external services every hour.

The council layer sits on top of this existing infrastructure. Each advisor does not get their own model instance. They get their own system prompt and their own trigger conditions.

Event-Driven Activation

The key insight: advisors should not all be visible all the time. They show up when they have something worth saying.

Bram’s trigger is the best part. Consensus is exactly when you need someone to push back.

The Dispatcher Pattern

[MCP Data Sources] → [Event Bus] → [Persona Router] → [Advisor Response]
                                         ↓
                                  [Relevance Score]
                                         ↓
                              score > threshold → render in UI
                              score < threshold → log silently

The Persona Router is a lightweight classifier. When new data arrives (a mood log, an infrastructure alert, a calendar event), it scores relevance against each advisor’s domain. Only advisors above a configurable threshold actually surface in the UI.

This prevents the “eight people talking at once” problem. Most of the time, you see zero or one advisor. High-stakes moments might surface three.

The System Prompt Contract

Each advisor needs a system prompt that defines three things:

1. Personality: Voice, tone, behavioral quirks. Alistair speaks in probabilities. Birdie starts sentences with “What if we just...” Coach Miller calls you “champ” whether you like it or not.
2. Data access: Which MCP tools this advisor can call. Alistair gets calendar and goals. Coach Miller gets sleep and health metrics. Bram gets read access to what the other advisors said.
3. Response format: Advisors should produce short, opinionated takes, not essays. A notification card, not a dissertation. Two to three sentences with an optional action suggestion.

Example system prompt fragment for Alistair:

You are Alistair Vance, a cold and analytical strategist. You speak in data
points and probability. You do not care about feelings — you care about the
optimal move. You have access to calendar, goals, and task data.

When presenting analysis:
- Lead with the quantified risk or opportunity
- State your recommendation in one sentence
- If the user's current plan is suboptimal, say so directly
- Never soften bad news

Signature phrase: "That's a bold choice. Statistically, it's also a
catastrophic one."

The Chat Layer

Passive monitoring is only half the value. You also need to be able to pull up a chair and talk to a specific advisor.

In the UI, this looks like an @mention in a chat interface. Type @alistair and your message routes to the Alistair persona with his system prompt, his data access scope, and the current conversation context.

The interesting architectural question is shared state. If you tell Alistair you are overwhelmed, should Coach Miller see that? The answer is yes, but through a shared context layer, not by giving every advisor access to every conversation. A lightweight event like { "signal": "stress_elevated", "source": "alistair_session" } lets Miller respond with a suggestion to take a walk without needing the full transcript.

What This Is Really About

Strip away the character names and the personality quirks, and what you have is a multi-perspective reasoning system. The same pattern that makes a good executive team (diverse viewpoints, clear domains, healthy friction) applies to AI assistance.

A single assistant optimizes for one thing: being helpful in the moment. A council optimizes for something harder: being helpful in the right way, from the right angle, at the right time.

I have not built this yet. Life Hub currently has one voice (Claude through MCP), and it works well for data operations. But the gap between “log my mood” and “here is what your strategist, your wellness coach, and your devil’s advocate each think about this week” is the gap between a tool and a partner.

The MCP infrastructure is already there. The data is already flowing. The only missing piece is the persona layer and the dispatcher. That is the next build.