DEEPINSIGHT: How I Built a "McKinsey Consultant in Your Pocket" with n8n + DeepSeek V3

Abstract:
DEEPINSIGHT generates ultra-detailed, 3000-4000 word strategic audit reports – complete with financial models, SWOT analyses, and persona cards – in minutes, just by inputting a product name and region. Here’s a full breakdown of how I built it on MuleRun.


1. The Problem DEEPINSIGHT Solves
Accessing professional market strategy consulting (like McKinsey or BCG) is incredibly expensive and time-consuming. For startups and solopreneurs, getting a comprehensive report that includes TAM/SAM/SOM estimations, competitive analysis, and 5-year financial projections within hours is almost impossible.
DEEPINSIGHT aims to automate and democratize this process.

2. How I Built It (The Build & Tech Stack)
My core architecture is lean and powerful, orchestrated by n8n:

  • :high_voltage: The Brain (Core Engine): DeepSeek V3 (accessed via SiliconFlow API). I chose V3 for its excellent cost-performance ratio in long-form text generation and complex reasoning.

  • :hammer_and_wrench: The Orchestrator (n8n Workflow):

    • Node 1: Form Trigger: Utilizes n8n’s native form to collect Product Name, Region, and Features.

    • Node 2: HTTP Request: This is where the magic happens. I configured an incredibly detailed System Prompt (over 1000 tokens of instruction) that forces the AI to act as a “McKinsey Principal Consultant,” demanding strictly structured outputs (tables, scorecards, KPIs).

    • Node 3: Code Node: A simple JavaScript snippet to process the JSON response from the API, ensuring stable and clean output formatting.

    • Node 4: Respond: Delivers the fully rendered, long-form report directly back to the user.

3. Key Decisions & “Secret Sauce” (What I Learned)
Throughout the build, I discovered that Prompt Engineering was the ultimate game-changer.
As you can see in my n8n node, I didn’t just ask for “a report.” Instead, I enforced strict requirements:

  • :cross_mark: No vague descriptions; specific data source citations like “according to Statista 2024” or “IBISWorld reports” are mandatory.

  • :bar_chart: Must generate ASCII-style visual scorecards and data boxes.

  • :money_bag: Must include concrete Break-even calculations and CAC/LTV ratios.

  • :bust_in_silhouette: Must create detailed persona cards with first-person quotes.

4. What Worked and What Didn’t (My Wins & Fails)

  • :white_check_mark: What Worked: Setting DeepSeek V3’s temperature to 0.3. For strategic reports, too much creativity can lead to fabricated data. A low temperature ensures logical coherence and data integrity.

  • :cross_mark: What Didn’t Work: Initially trying to accomplish everything in a single, overly complex prompt often led to truncated outputs. I later optimized max_tokens to 12000 and refined the prompt structure for full, complete generation in one go.

5. Conclusion
This Agent proves that with well-structured prompt architecture and a flexible tool like n8n, we can transform expensive professional services into reproducible APIs.

1 Like

Hi @liang_caoqw1, thanks for sharing your showcase!

A couple of suggestions: 1) Add “MuleRun” to your blog title so people can see how it’s related to this community. 2) Drop your agent link into the blog to drive more users to try it.