People always ask me whether they should use ChatGPT, Claude, or Gemini. Which one? And for what?

It's a fair question, and the answers are nuanced. But it's fundamentally the wrong one.

We're so obsessed with model selection - benchmarks, updates, features- that we miss the elephant in the room.

The real secret isn't choosing between Claude 4.0 Opus or ChatGPT o3. It's context engineering.

Context Engineering: What Really Matters

Context engineering means giving your AI the precise information it needs to deliver institutional-grade insights.

Think of AI as the brightest analyst you've ever hired with zero institutional memory. That straight-out-of-MIT math prodigy that can do everything you could ever dream of better than any other analyst at your firm, but you have to hold his hand.

The heavy hitters in Silicon Valley are talking about this - see below from Andrej Karpathy, one of the founding members of OpenAI.

Andrej Karpathy @karpathy +1 for "context engineering" over "prompt engineering". People associate prompts with short task descriptions you'd give an LLM in your day-to-day use. When in every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window https://twitter.com/tobi/status/1935533422589399127 tobi lutke @tobi I really like the term “context engineering” over prompt engineering. It describes the core skill better: the art of providing all the context for the task to be plausibly solvable by the LLM. 10:54 AM • Jun 25, 2025 2076 Retweets 13996 Likes Read 532 replies

​

​

For investors, this means providing:

• What type of investor are you?
• What is your coverage universe?
• What is your investment philosophy?
• How are decisions taken at your firm, and what specific data points help guide those decisions?
• How do you conduct your research and analysis?

Without context, your LLM is really just a very smart guesser. With the right context, you can get a lot more insights that help you make you make better decisions. And that in turn equals higher returns.

The 5-Block Context Packet Framework (ROOCS)

After writing thousands of prompts myself, I've boiled context engineering down to five blocks that you should include in every prompt:

1. ROLE: Define precisely who your AI is simulating- a generalist distressed debt investor, a consumer megafund private equity associate, a TMT sell-side equity research analyst, etc.
2. OBJECTIVE: Clearly state the outcome you want in one to two concise sentences.
3. OUTPUT: State in as much detail as possible what you want the output to be. Bulleted lists work best.
4. CONTEXT: Include:
   • Audience: Who exactly is this analysis for (your IC, a PM / managing partner that wants high-level insights, your junior associate that is deep in the weeds)?
   • Voice & Tone: Authoritative, critical, supportive-tailor this based on audience.
   • Length Target: Specify how concise or detailed you want it. For most prompts, I prefer lengthy explanations (instead of crisp bullets), which result in 1,000-10,000 word outputs.
   • Must-Use Inputs: Essential data, documents, and metrics that must underpin the analysis.
   • Constraints: Guardrails like "no forward-looking statements," precise sourcing requirements, or specific formatting rules.
5. STEPS:
   1. Gap check - list any information or context that is still missing. Ask me concise questions until the gaps are filled
   2. Plan - outline a logical structure or bullet agenda for the piece. Wait for my approval
   3. Draft - write the first version following the approved plan
   4. Review - pause and ask me for feedback on clarity, tone, completeness, and general direction
   5. Revise - iteratively improve the draft with your comments

It's almost the same principle as sending comments to your team- you iterate until you have something that works. The difference is the AI works much faster and isn't afraid to ask you for clarification.

This is how you get from the AI slob we've been accustomed to in the past years to institutional-grade materials that are actually helpful.

Example 1: Turning an S-1 into a Three-Page IC Memo

Figma just launched its IPO roadshow- big news for capital markets. For your analysts, this implies reading their 200-page S-1 and briefing their PM ahead of the investment committee.

Here’s how you deploy context engineering:

See how it works.

Using ChatGPT o3, I attached the PDF of Figma's S-1 + activated "Deep Research".

Then you must answer ChatGPT's questions so that it has appropriate context:

And now ChatGPT will take care of the research and drafting on your behalf.

Example 2: Finding Blind Spots in Your Analysis

Suppose you've drafted your memo and now want an independent voice to challenge your assumptions. Here’s how to leverage the AI as your internal risk manager:

Having a second run at the LLM with a "find my blind spots" prompt - loaded with the prior context, its last output, and your new inputs - will produce far more critical, zero‑fluff insight

Here's the full chain of analysis: https://chatgpt.com/share/687f1478-39e4-8007-892b-366bce3bd02e​

When Not to Use Context

Is detailed context always necessary? No. Context engineering shines brightest in complex, high-stakes analysis: situations requiring nuanced understanding, careful reasoning, or risk assessments. But when you need quick, factual, straightforward responses, go ahead and fire off a simple, context-free prompt.

• Quick text fixes (spelling, grammar, rewording, reformatting)
• Simple math
• Small data grabs (pull x data point for y company)

For these tasks, context would be overkill-quick prompts suffice.

The Bottom Line

While everyone debates models and each one has its merits, the most important thing is on building context systems.

I've tested identical analyses across ChatGPT, Claude, and Gemini. With proper context, they all deliver institutional-grade insights. Without it, they all fail.

The difference isn't the model. It's what you feed it. That's how you get your top-bucket AI analyst.

Happy context engineering.

​