Building the Advisory Board MCP: Transforming Notion Pages into Strategic War-Rooms with the 7 Hats Framework

This is a submission for the Notion MCP Challenge

What I Built

I built the Notion Advisory Board MCP, an automated strategic review system.

Often, project proposals and technical specs get rubber-stamped because everyone just wants the meeting to end. I wanted to build an AI workflow that acts as an unforgiving, data-driven Advisory Board to do the stress-testing we're all too tired to do ourselves. By simply pasting a Notion page link into Claude, this MCP server orchestrates a full 7 Hats Analysis (White, Red, Black, Yellow, Green, Blue, and Technical hats) because why have one opinion when you can artificially generate seven?

Instead of just hallucinating risks like a standard LLM, the server actively queries a Lessons Learned database in your Notion workspace to dig up real past company failures. It then forces the Black Hat and Technical Hat to weaponize those historical blunders against the current proposal. Finally, it writes the entire debate and a structured Pro/Con recommendation directly into the original Notion document, so you can watch your beautiful idea get methodically dismantled in real-time.

Video Demo

Youtube
Notion - Sample Proposal & Analysis
Notion - Sample Lessons Learned DB

Show us the code

Github

How I Used Notion MCP

This project heavily leverages the @modelcontextprotocol/sdk and the @notionhq/client. Integrating Notion via MCP unlocked the ability to bridge the gap between working documents (the proposal) and company memory (the database).

Here is how the MCP integration powers the workflow:

1. Tool-Guided Orchestration (analyze_proposal): The first tool reads the blocks of the target Notion page. To ensure the user doesn't have to type a massive prompt every time, this tool automatically appends hidden "System Instructions" to the returned text, forcing Claude to immediately run the 7 Hats framework without asking for permission.

2. Prompt Injection Defense: Because we append system instructions to user-generated Notion text, the LLM initially flagged the documents for prompt injection! I utilized XML tags ( vs ) to safely parse the Notion data via MCP.

3. Data-Backed Context (get_historical_lessons): This tool uses Notion's databases.query endpoint. Claude extracts keywords from the proposal and searches the workspace's "Company Brain" database, bringing past post-mortems directly into the context window so the AI's critique is grounded in reality.

4. Native UI Generation (post_opinion & post_summary): Instead of keeping the analysis isolated in the chat UI, these tools use Notion's blocks.children.append API to write directly to the proposal page. I mapped specific Agent Personas to Notion's native Callout block colors/emojis, creating a cascaded report that lives permanently alongside the original spec.

What Can Be Improved

While the Advisory Board is currently excellent at tearing down bad ideas, there is always room for more sophisticated ruthlessness:

• Multi-Database Context: Right now, the Cynic only queries the "Lessons Learned" database. Future iterations could query a live Budgets or Team Capacity database to automatically flag if a proposal is too expensive or if the engineering team is already at 120% capacity.
• True Human-in-the-Loop Orchestration: Instead of dumping the entire 7 Hats analysis and summary at once, the MCP server could pause after the debate, append a "Waiting for PM Approval" checkbox to the Notion page, and poll that block. The AI would only generate the final roadmap once a human physically checks the box.
• Parallel Processing: Currently, Claude processes all 7 hats in a single massive chain of thought. Offloading this to a swarm of smaller, specialized, parallel agent processes would speed up the response time significantly (and make the board meetings feel delightfully chaotic).