Mind Council

Inspiration

We've all been there — a decision that feels too big to get wrong, and too complex to think through alone.

One of our teammates spent three weeks trying to decide between a startup offer and a master's program. She asked her parents, her professors, her friends. She made four different pros-and-cons lists. She researched salary data, spoke to alumni, and filled an entire notebook. And yet — she still felt paralyzed.

The problem wasn't a lack of information. It was a lack of structured reasoning.

That moment became the seed for MindCouncil. We were inspired by the idea that every great decision-maker in history didn't just think harder — they sought diverse perspectives. A board of advisors. A council of voices. People who would challenge their assumptions, surface their blind spots, and model what each path actually looked like.

We asked: What if anyone could have that council — powered by AI?

"The goal of MindCouncil is not to tell you what to decide. It is to make sure you've thought about everything you should — before you do."

What it does

MindCouncil is an AI-powered life decision simulator and multi-advisor reasoning engine. It combines two capabilities into one seamless experience:

1. Decision Simulator (DecisionLens core)

Users bring a real decision — grad school vs. job, startup vs. employment, city A vs. city B. MindCouncil models each path across 5 dimensions.

2. The Council of Five Advisors

After the simulation, five AI personas each give their honest, unfiltered take:

Advisor	Lens	Focuses on
🔵 The Pragmatic Parent	Security & stability	Financial risk, long-term safety
🟡 The Risk-Taking Entrepreneur	Growth & opportunity	Upside potential, bold moves
🟢 The Cautious Academic	Evidence & depth	Research, credentials, long-term learning
🔴 The Values-First Counselor	Meaning & identity	What matters to you, not the world
🟣 The Data-Driven Analyst	Numbers & patterns	Market data, acceptance rates, salary ranges

3. Hidden Tradeoff Surfacer

MindCouncil proactively identifies 3 tradeoffs the user hasn't mentioned — the things you forgot to consider. This is the feature users consistently say changes their decision process.

4. Clarity Report

Every session ends with a shareable Clarity Report: scenario comparison, council summaries, uncertainty map, and one concrete next step — never a final answer.

How we built it

We built MindCouncil as a full-stack AI web application over 72 hours during the build window.

Architecture

User Input → Diagnostic Engine → Tradeoff Analysis → Scenario Simulation → Council Generation → Clarity Report

Tech Stack

Frontend

React + Vite + TailwindCSS
shadcn/ui component library
Recharts for radar/scenario visualization
Streaming API responses for real-time AI output

Backend

FastAPI (Python) — async API endpoints
Supabase (PostgreSQL) — session and decision storage
Supabase Auth — optional Google OAuth

AI Layer

Claude API (claude-sonnet-4-6) — core reasoning engine
Multi-prompt orchestration pipeline:
1. Diagnostic prompt — Socratic questioning to surface user values and constraints
2. Tradeoff prompt — identifies 3 gaps in the user's stated considerations
3. Scenario prompt — structured JSON output with dimensional scores + narratives
4. Council prompt — 5 parallel API calls, each with a distinct advisor persona system prompt
5. Report prompt — plain-language Clarity Report generation

Deployment

Frontend: Vercel
Backend: Railway
Storage: Supabase Storage (PDF export)

Prompt Engineering Approach

The most critical engineering challenge was designing prompts that produce structured reasoning, not content generation. We enforced JSON output schemas for all simulation responses and built a validation layer using Pydantic to catch and re-prompt malformed outputs.

Each advisor persona was carefully crafted with a distinct voice, priorities, and blind spots — the Entrepreneur persona, for example, is explicitly instructed to underweight financial risk in ways the Analyst would flag.

Challenges we ran into

1. Making AI reason, not just respond

The hardest problem wasn't technical — it was philosophical. Every time we tested an early prototype, the AI wanted to answer the question. Getting Claude to consistently hold back from giving a recommendation while still being genuinely useful required significant prompt iteration. We went through 14 versions of the diagnostic prompt before it felt truly Socratic.

2. Structured output reliability

We needed scenario scores as valid JSON — dimensional scores, confidence ranges, uncertainty flags — reliably, every call. LLMs are probabilistic; they don't always return clean JSON. We built a two-layer fallback system: Pydantic validation on first parse, and an automatic re-prompt with a stricter schema on failure.

3. The 5 advisor voices diverging too much

Early versions of the Council had advisors that were so different they confused users rather than helping them. The Entrepreneur said "do it," the Parent said "never," and users felt more paralyzed than before. We recalibrated — each advisor now engages with the same tradeoffs but through a different lens, so the council feels like a conversation, not a shouting match.

4. Streaming with parallel API calls

The Council of Five requires 5 simultaneous Claude API calls. Streaming each one to the frontend while preserving the order of advisors and handling partial failures gracefully was a real engineering challenge. We solved this with a Promise.allSettled() architecture and individual streaming buffers per advisor.

5. Representing uncertainty honestly

Confidence ranges like $7 \pm 2$ sound simple but are surprisingly hard to communicate without either alarming users or making scores feel meaningless. We iterated on the Clarity Report visualization three times before landing on a design that made uncertainty feel informative rather than worrying.

Accomplishments that we're proud of

The Socratic Engine works. Users consistently report that the diagnostic questions surface priorities they hadn't consciously articulated. One beta tester said: "It asked me something I'd never thought to ask myself."
The Council of Five feels genuinely diverse. Each advisor has a distinct, consistent voice across sessions. The Cautious Academic sounds nothing like the Risk-Taking Entrepreneur — and that contrast is where the value lives.
We shipped a working, deployed product in 72 hours. Not a mockup. Not a demo. A real application with streaming AI responses, a radar chart, scenario comparison, and a shareable Clarity Report.
The Responsible AI layer is real, not decorative. MindCouncil has a hard-coded constraint: it never states a recommendation. The final screen always reads "The right choice is yours — here is what you now know to make it." We're proud that this isn't just a disclaimer — it's baked into the system prompt at the architecture level.
We validated with real users. We ran 12 sessions with actual undergraduate students facing real decisions during the build window. Average self-reported confidence increase: 68%. Zero users felt the AI told them what to do. All 12 said they felt clearer after.

What we learned

The most important lesson: AI that asks better questions is more valuable than AI that gives better answers.

We came in expecting to build a recommendation system. We ended up building a reasoning companion. That shift in framing changed everything — the prompts, the UX, the success metrics, the ethical design.

On prompt engineering: Persona consistency at scale is hard. A persona isn't just a name and a one-line description — it's a coherent set of priorities, blindspots, communication style, and values that has to hold across arbitrary inputs. We now think of persona prompts like character sheets in fiction writing.

On uncertainty communication: Users trust AI more when it admits what it doesn't know. Counter-intuitively, showing $7 \pm 2$ instead of $7$ made users more confident in the tool, not less. Honesty builds trust.

On responsible AI: Human-in-the-loop isn't a checkbox — it's a design philosophy. Every screen, every CTA, every piece of copy in MindCouncil was written to reinforce that the user is the decision-maker. That required discipline, because the easy thing is always to let the AI feel more authoritative than it should.

On team dynamics in a hackathon: Split responsibilities by thinking style, not just skill. Our best decision was having our "product thinker" own the prompt engineering — not the engineers. The prompts are the product.

What's next for Mind Council

MindCouncil is solving a real problem for tens of millions of people. Here is the roadmap:

Near-term (0–3 months)

Mobile app (React Native) — most life decisions happen on the go
Decision templates for common scenarios: career, grad school, city relocation, startup vs. employment
Real data integration — live salary data (BLS API), cost-of-living (Numbeo), grad school acceptance rates
Shareable Clarity Reports — PDF export and public link for sharing with mentors

Medium-term (3–12 months)

University career center partnerships — white-label MindCouncil for campus deployment
Longitudinal tracking — did your scenario play out the way the model predicted? Close the feedback loop
Group decision mode — startup co-founders making product or strategic decisions together
Decision history and pattern recognition — "You consistently prioritize security over growth. Here's how that's shaped your past decisions."

Long-term vision

We believe decision intelligence is the next frontier in AI assistants. Not AI that acts for you — AI that helps you think better. MindCouncil's goal is to become the reasoning layer that sits behind every major life decision, the way a calculator sits behind every financial one.

The strongest projects turn from hackathon submissions into products people actually use. MindCouncil is built to be one of them.