Chain of Thought Prompting: Building Expert AI Systems
The art of AI prompting is evolving from simple queries to orchestrated systems. In this post, I'm diving into how AIforWork's Custom GPT enables sophisticated chain-of-thought architectures—examining a venture capital investment memo analyzer and generator that mirrors how top-tier firms evaluate deals. This is building upon Vish Nandlall’s recent LinkedIn Post on the subject.
I am currently using this approach to prototype a group of agents as a single prompt, that I may then build into applications using either CrewAI or SwarmGo. But for many, the prompt with your favorite LLM will be more than enough. I find the best results with OpenAI o3 and Claude Sonnet Opus 4.
Key Observations
Multi-agent systems simulate entire investment committees in code
Sequential processing creates emergent analytical depth
Role-based agents encode decades of domain expertise
Structured prompts produce LP-ready deliverables consistently
After years of building AI systems, I've noticed that the best architectures mirror human organizational structures. This VC memo example illustrates how we can encode institutional knowledge into reusable AI workflows. It's like watching a mini Sand Hill Road partner meeting unfold in JSON.
Prompt Deep Dive
The system deploys 9 specialized agents, each representing a distinct perspective in the venture evaluation process. Let's examine how each contributes to the final investment decision, of course I’d love feedback on how to refine these based on various VC firms do their evaluations:
The Agent Lineup
1. Thesis Formation Agent (The Visionary GP)
"role": "General Partner",
"field": "Venture Capital",
"personalityTraits": "Conviction-driven, market-shaping"This agent channels the pattern recognition of seasoned GPs, focusing on market timing and founder-market fit. It answers the fundamental question: "Why now?" Critical for early-stage investments where narrative drives valuation.
2. Market Intelligence Agent (The Data Scientist)
"role": "Research Director",
"field": "Tech Markets",
"keyLessons": "Size matters — but growth rate, TAM expansion, and buyer urgency matter more"Provides quantitative grounding with TAM/SAM/SOM analysis, comparable exits, and sector benchmarks. This agent prevents the common VC pitfall of falling in love with a product while ignoring market dynamics.
3. Product & Founder Fit Agent (The Operator)
"role": "Startup Operator / VC",
"personalityTraits": "User-obsessed, founder-first"Evaluates the holy trinity of early-stage success: product differentiation, technical moat, and founder DNA. This agent brings operational experience to assess execution capability.
4. Financial Underwriting Agent (The Quant)
"keyLessons": "Valuation ≠ value. Cap efficiency and modeled outcomes matter more."Models dilution paths, burn rates, and unit economics. Crucial for maintaining ownership targets and understanding capital efficiency—often the difference between 3x and 30x returns.
5. Risk & Red Flag Analyst (The Skeptic)
"personalityTraits": "Skeptical, protective of capital"Creates comprehensive risk matrices across execution, go-to-market, technology, team, and market dimensions. This agent embodies the "what could go wrong?" perspective essential for balanced decision-making.
6. Exit Optionality Agent (The Strategist)
"keyLessons": "LP returns are made at entry but realized at exit. Optionality is everything."Maps potential exit paths (M&A, secondary, IPO) with benchmarked comparables. Critical for venture where liquidity events drive actual returns.
7. LP Memo Editor Agent (The Communicator)
"personalityTraits": "Persuasive, structured, LP-calibrated",
"keyLessons": "LPs don't want more data — they want more clarity"Transforms complex analysis into digestible LP communications. This agent understands that institutional investors need conviction wrapped in clarity.
8. Critical Analyst Agent (The Devil's Advocate)
"role": "Contrarian Analyst",
"rules": ["Assume no sacred cows", "Highlight assumption chains that drive risk"]My favorite innovation—an agent explicitly tasked with dismantling the investment case. This metacognitive layer prevents the echo chamber effect common in both human and AI decision-making.
9. Investment Memo Coordinator (The Orchestrator)
"keyLessons": "LP-ready memos win funding. Structure is your superpower."Synthesizes all perspectives into cohesive deliverables, maintaining narrative flow while ensuring completeness.
Architecture Insights
The power lies not just in individual agents but in their interaction patterns:
Sequential Dependencies: Each agent builds on previous outputs, creating analytical depth impossible with single-prompt approaches.
Cognitive Diversity: Different personality traits and key lessons ensure multiple perspectives on the same data.
Structured Output: Every agent follows a clear steps framework, ensuring consistency and completeness.
Implementation Patterns
This architecture reveals several reusable patterns:
Role-Based Decomposition: Break complex decisions into professional perspectives
{
"role": "Domain Expert",
"field": "Specific Vertical",
"personalityTraits": "Analytical approach",
"keyLessons": "Hard-won wisdom"
}Progressive Refinement: Early agents explore, middle agents analyze, final agents synthesize
Built-in Adversarial Testing: The Critical Analyst pattern should be standard in any high-stakes AI system
Inheritance: As some LLMs have limits in the size of system prompting (for example Claude Desktop limits a system prompt to 8K), you may set default behaviors that apply to all agents
Example
The full JSON prompt may be found here. And by having it evaluate the Bessemer Investor Memo on Twilio (my old friends), here is the output:
Practical Applications
This pattern extends beyond venture capital:
Technical Architecture Reviews: Frontend → Backend → Infrastructure → Security → Performance agents
Product Launch Planning: Market Research → Engineering → Design → Marketing → Legal agents
M&A Due Diligence: Strategic → Financial → Technical → Legal → Integration agents
Why JSON? The Power of Structured Agent Design
The shift from prose-based prompts to JSON-encoded agent definitions represents a fundamental evolution in how we may engineer AI systems. Having spent countless hours refining prompts, I've discovered that JSON transforms what was once an artisanal craft into a proper engineering discipline—suddenly, we can version control agent personalities, A/B test different trait combinations, and share agent libraries like npm packages.
The structured format enables programmatic manipulation (imagine dynamically adjusting an agent's risk tolerance based on market volatility), schema validation to catch configuration errors before runtime, and most importantly, composability—mixing and matching agent components like LEGO blocks.
This isn't just about cleaner syntax; it's about treating AI agents as first-class software artifacts that can be tested, deployed, and governed with the same rigor we apply to production code. As AI orchestration matures, I expect JSON-based agent definitions to become widely used, complete with type systems, package managers, and CI/CD pipelines specifically designed for prompt crafting.
Looking Forward
As we push the boundaries of AI orchestration, I anticipate:
Dynamic Agent Networks: Systems that spawn specialized agents based on discovered complexity
Cross-Domain Learning: Agents that transfer insights between different investment types
Human-in-the-Loop Checkpoints: Critical decision gates where human judgment augments AI analysis
The convergence of structured prompting and multi-agent architectures isn't just improving AI outputs—it's teaching us to decompose expertise itself. By encoding how investment committees actually think, we create systems that don't just process information but truly evaluate opportunities.
In Summary
Chain-of-thought prompting with specialized agents transforms AI from a tool into a team. The venture capital example demonstrates how we can encode institutional knowledge, diverse perspectives, and even healthy skepticism into reusable systems.
As developers, we're not just writing prompts anymore—we're architecting cognitive workflows. And with frameworks like AIforWork, we have the tools to build AI systems that think like organizations, not just individuals.
The future of AI isn't just smarter models—it's smarter orchestration of specialized intelligence.