Build an AI Perspective Council with ChatGPT, Python and LangGraph
PO Prompts #003 | Think the Unthinkable
Project Description
What if every employee had 100 AI assistants—but none of them was allowed to do the employee's work?
This deliberate provocation challenges the conventional assumption that the primary purpose of AI assistants is to automate tasks.
Instead, imagine AI assistants acting as independent thinking partners, examining a problem from different perspectives, challenging assumptions, identifying overlooked risks and discovering unexpected opportunities.
The proposed innovation: An AI Perspective Council—a multi-perspective decision-support system that helps people make better-informed decisions without surrendering human judgment.
The project begins with five AI perspectives and can later expand to a larger, dynamically selected council.
1. Project Objective
Design, develop and evaluate an AI-powered decision-support application in which multiple specialized AI perspectives examine the same problem, exchange critiques and produce a consolidated briefing for a human decision-maker.
The system should:
Challenge conventional assumptions.
Explore alternative interpretations of problems.
Identify opportunities and potential consequences.
Surface disagreements and missing evidence.
Generate unconventional but actionable options.
Support transparent, accountable human decisions.
Expected outcome: A working AI Perspective Council prototype that can be tested against conventional single-assistant AI decision support.
2. Lateral Thinking Foundation
Conventional Thinking
One employee encounters a problem, asks an AI assistant for recommendations and receives an answer.
Adding more assistants might simply mean automating more tasks.
PO Provocation — Exaggeration
PO, every employee has 100 AI assistants, but none can perform the employee's tasks.
This is a deliberate thinking provocation, not a literal implementation requirement.
MOVE — Extract a Principle
If assistants cannot execute tasks, what useful roles remain?
They could challenge assumptions, examine consequences, uncover alternatives, question evidence and reveal opportunities.
The underlying principle is that AI assistance can increase the diversity of thinking rather than merely the volume of work completed.
New Direction
Create an AI council in which specialized perspectives examine a problem independently and then challenge one another.
The goal is not to manufacture artificial agreement.
It is to expose important differences in reasoning and help humans discover options they might otherwise overlook.
GENERATE
Use generative AI to develop multiple perspectives, explore emerging alternatives and refine unconventional ideas.
JUDGE
Assess novelty, relevance, feasibility, evidence quality, cost, disagreement and risks.
ACCOMPLISH
Build and evaluate a five-perspective AI Council MVP.
Distinction: The exaggeration and movement techniques are inspired by Edward de Bono's lateral-thinking methods. The AI Perspective Council is a proposed contemporary application, not a technique attributed to de Bono.
3. Recommended Technology Stack
| Component | Recommended Tool | Purpose |
|---|---|---|
| AI models | OpenAI GPT models | Generate and critique alternative perspectives |
| No-code prototype | ChatGPT Custom GPT | Test council-style prompting |
| Programming language | Python | Implement application logic |
| AI integration | OpenAI API | Invoke specialized AI perspectives |
| Workflow orchestration | LangGraph | Coordinate independent analysis, critique and synthesis |
| User interface | Streamlit | Build an interactive decision-support dashboard |
| Structured outputs | Pydantic | Validate perspective responses and decision briefs |
| Data storage | SQLite | Store scenarios, analyses and evaluation results |
| Data analysis | pandas | Compare outcomes across experiments |
| Visualization | Plotly | Display scores, disagreements and evaluation results |
| Testing | pytest | Test workflow behavior and output validation |
| Deployment | Streamlit Community Cloud | Host a non-sensitive demonstration |
Recommended Development Strategy
Stage A — No-code: Simulate five perspectives using a Custom GPT or structured ChatGPT prompt.
Stage B — Low-code: Build a Streamlit interface connected to the OpenAI API.
Stage C — Orchestrated application: Introduce LangGraph for independent analysis, controlled critique, synthesis and review.
A multi-agent framework is optional for the first prototype. Multiple perspectives can initially be implemented using separate model calls with distinct instructions.
4. Proposed System Architecture
Human Decision Problem → Context Validation → Perspective Selection → Independent AI Analysis → Cross-Perspective Critique → Evidence and Risk Review → Decision Brief → Human Judgment
Five Core AI Perspectives
1. The Customer Advocate
Examines customer needs, usability, expectations and possible dissatisfaction.
2. The Critical Challenger
Questions assumptions, identifies weak reasoning and challenges proposed solutions.
3. The Evidence Researcher
Identifies evidence requirements, unsupported claims, uncertainty and information gaps.
4. The Risk Analyst
Examines operational, financial, ethical, regulatory and implementation risks.
5. The Innovation Explorer
Uses deliberate provocations and movement techniques to discover unconventional alternatives.
Council Coordinator
A separate coordinating workflow collects the perspectives, highlights disagreements and prepares a consolidated decision brief.
The coordinator must not silently erase minority views or convert speculation into established evidence.
Human Decision-Maker
The user reviews the council's findings, requests clarification, compares alternatives and makes the final decision.
5. Reusable Prototype Prompt
Create a Custom GPT called AI Perspective Council using the following instructions:
You are an AI Perspective Council designed to improve human decision-making through multiple independent perspectives.
When the user presents a decision problem:
Clarify the decision, objectives, constraints and available evidence.
Simulate five perspectives: Customer Advocate, Critical Challenger, Evidence Researcher, Risk Analyst and Innovation Explorer.
Have each perspective independently examine the problem and identify important considerations.
Require the Innovation Explorer to apply a deliberate lateral-thinking provocation and movement technique.
Identify agreements, disagreements and conflicting assumptions.
Generate at least three distinct decision alternatives, including one non-obvious option.
Evaluate alternatives against user-defined criteria.
Identify missing evidence and suggest practical validation experiments.
Prepare a concise decision brief with recommendations, uncertainties and unresolved disagreements.
Leave the final decision to the human user.
Never fabricate evidence, external research, expert identities or independent verification.
Clearly distinguish facts, assumptions, AI-generated hypotheses and recommendations.
Do not force consensus when meaningful disagreements remain.
Implementation note: A Custom GPT can simulate multiple perspectives, but this does not establish that the analyses are truly independent. The application version should use separate model calls and controlled information sharing to better test perspective diversity.
6. Minimum Viable Product (MVP)
MVP Objective
Build a functional web application where a user enters a workplace decision and receives a structured analysis from five AI perspectives.
Essential Features
Feature 1 — Decision Input
The user enters:
Decision question
Business context
Desired outcome
Constraints
Evaluation criteria
Feature 2 — Five-Perspective Analysis
The application invokes five separately instructed AI perspectives and displays their findings.
Feature 3 — Assumption Challenger
The system identifies at least three assumptions that may be limiting the decision.
Feature 4 — Lateral Thinking Explorer
The Innovation Explorer applies one explicit provocation method and one movement technique to develop a non-obvious alternative.
Feature 5 — Council Discussion
Perspectives critique selected claims or proposals from other perspectives.
Feature 6 — Decision Comparison
The system compares at least three alternatives using a transparent scoring rubric.
Feature 7 — Decision Brief
The application generates a report containing:
Problem definition
Important assumptions
Five perspective summaries
Agreements and disagreements
Alternative solutions
Evidence gaps
Risk considerations
Suggested experiments
Human decision checklist
Feature 8 — Export
Allow the user to export the briefing as Markdown or PDF.
Example Use Case
Decision: Should a small company replace its customer support team with an AI chatbot?
The five perspectives might respond as follows:
| Perspective | Key Question |
|---|---|
| Customer Advocate | What happens when customers need empathy or escalation? |
| Critical Challenger | Are we assuming that automation requires replacement? |
| Evidence Researcher | What evidence supports the expected savings and service quality? |
| Risk Analyst | What happens when the chatbot gives incorrect or sensitive advice? |
| Innovation Explorer | What if human agents handled only the moments where human judgment matters most? |
Possible innovation: Instead of full replacement, design an AI-supported customer service system where routine queries are automated and human staff specialize in complex, emotional or high-consequence interactions.
This is a hypothesis to test, not a predetermined recommendation.
7. Evaluation Experiment
Research Question
Does a multi-perspective AI Council produce better-supported and more innovative decisions than a conventional single AI assistant?
Experimental Design
Prepare 10–20 realistic decision scenarios spanning business, education, technology and workplace operations.
Compare three conditions:
A — Human-only: Participants analyze the problem without AI.
B — Single AI Assistant: Participants receive a conventional AI-generated recommendation.
C — AI Perspective Council: Participants receive the council's structured analysis and disagreements.
Use the same problem information and evaluation criteria across conditions. Where feasible, randomize or counterbalance scenarios and have reviewers assess anonymized decision outputs.
Evaluation Metrics
| Metric | Measurement |
|---|---|
| Decision quality | Independent reviewer rubric |
| Perspective diversity | Number of substantively distinct considerations |
| Assumption discovery | Relevant hidden assumptions identified |
| Innovation quality | Novelty and usefulness of alternatives |
| Evidence quality | Accuracy and traceability of claims |
| Risk coverage | Relevant risks identified |
| Human understanding | Participant comprehension of trade-offs |
| Decision confidence | Self-reported confidence and calibration |
| Efficiency | Time, latency and API cost |
| Usability | Participant feedback |
Important Evaluation Safeguards
More perspectives do not automatically mean better decisions.
Multiple AI perspectives can share biases, repeat the same errors or create a misleading appearance of independent agreement.
The experiment should therefore measure meaningful diversity, factual reliability, user workload and costs—not merely the quantity of generated ideas.
A small pilot provides preliminary findings rather than proof of general effectiveness.
8. Suggested Development Roadmap
Phase 1 — Concept Validation
Duration: Approximately 2–3 days
Create a Custom GPT prototype and test five representative decision problems.
Deliverable: Working prompt-based council demonstration.
Phase 2 — MVP Development
Duration: Approximately 1–2 weeks
Build a Python and Streamlit application using the OpenAI API.
Implement independent perspective calls, structured responses, a comparison view and decision brief export.
Deliverable: Functional web prototype.
Phase 3 — Council Orchestration
Duration: Approximately 1 week
Introduce LangGraph if the workflow requires explicit stages, controlled critique and review checkpoints.
Deliverable: Coordinated multi-perspective workflow.
Phase 4 — Quality and Safety Testing
Duration: Approximately 1 week
Test prompt injection resistance, unsupported claims, output consistency, privacy protections, costs and error handling.
Deliverable: Test report and revised prototype.
Phase 5 — Comparative Evaluation
Duration: Approximately 2 weeks
Run decision experiments with representative users and independent reviewers.
Deliverable: Evaluation dataset and comparative analysis.
Phase 6 — Improvement and Expansion
Explore additional perspectives, adaptive perspective selection, verified document retrieval and domain-specific councils.
Deliverable: Enhanced prototype and implementation recommendations.
Durations are illustrative estimates for a small prototype, not guaranteed delivery commitments.
9. Expected Project Deliverables
AI Perspective Council Custom GPT prototype.
Reusable multi-perspective prompting framework.
Python and Streamlit web application.
Five specialized AI perspective modules.
Optional LangGraph orchestration workflow.
Structured decision-analysis templates.
Decision brief export functionality.
Evaluation dataset and scoring rubric.
Comparative pilot report.
Technical documentation and deployment guide.
10. Future Extensions
Once the MVP is evaluated, possible enhancements include:
Dynamic Perspective Selection: Choose the most relevant perspectives for each problem rather than always using the same five.
100-Perspective Library: Maintain a catalog of specialized perspectives, invoking only a useful subset for each decision.
Evidence-Grounded Council: Connect approved documents and verified information sources so perspectives can support claims with traceable evidence.
Human-AI Debate: Allow users to challenge individual perspectives and request counterarguments.
Decision Memory: With appropriate permissions, record decisions, assumptions, outcomes and lessons learned for future review.
Specialized Councils: Develop versions for AI engineering, education, entrepreneurship, product design and organizational governance.
11. What Makes This an AI Creativity Project?
Conventional AI assistance often follows a familiar pattern:
Ask → Generate → Select
The AI Perspective Council explores a different approach:
Disrupt → Move → Discover → Explore with AI → Judge → Accomplish
The project begins by exaggerating the number of AI assistants while removing their ability to execute tasks.
That provocation creates an unexpected question:
What else could AI assistants contribute besides doing work?
Through deliberate movement, we discover the possibility of AI systems designed to expand the diversity of human thinking.
The goal is not to replace human judgment with an artificial committee.
It is to help people recognize assumptions, understand disagreements, discover alternatives and make better-informed decisions.
Lateral thinking is the creative engine. AI is the exploration amplifier. Accomplishment is the destination.
Project status: Proposed innovation concept and technical blueprint. Prototype development and experimental validation remain future work.
PO Prompts — Think the Unthinkable.
AIYING — Accomplishing with AI.
