How to think about AI as a research partner, not a research tool

Most investors use Claude the same way they use a search engine.

Ask a question. Get an answer. Mvoe on

That’s fine. But it’s also why most people get mediocre results and assume that’s just how it worrks.

The thing that changed how I use AI in research wasn’t a better prompt. It was a different expectation for what the conversation should be..

A tool does what you ask. A partner pushes back.

When you treat Claude as a tool, you ask it to do things. Summarize this filing. List the competitors. Pull out the margin history.

You get clean outputs quickly. Useful, no question.

But you’re still doing all the thinking. Claude is just handling the mechanical parts faster than you could.

When you treat it as a partner, the dynamic shifts. You share your reasoning, not just your question. You explain what you’re trying to figure out, what you currently believe and why, and where you’re not sure you’re right.

Then you ask it to engage with that.

The quality of what comes back is in a different category.

Why the tool approach limits you

Think about working with a good analyst.

You wouldn’t hand them a company name and say “list the risks.” You’d walk them through your current thinking, explain what you’re seeing, tell them where you’re uncertain and ask them to push back.

The conversation would move back and forth. They’d challenge something. You’d either defend it or realize you couldn’t. They’d bring up an angle you hadn’t thought about. Your thinking would shift in real time.

That’s what research actually is. The list of risks is just a byproduct.

Most investors never have that conversation with Claude because they never set it up that way. They ask for the list and close the tab.

What this looks like when you do it differently

Instead of: “What are the main risks for this company?”

Try this instead: “I’m looking at this company because I think the market is underpricing their switching costs. My reasoning is that customers integrate the product into their billing workflow, which makes switching genuinely painful and expensive. I think the market is discounting this because churn spiked two years ago, but I believe that was a one-time pricing decision, not a structural problem. Here’s the relevant section from the 10-K. Where is my logic most likely to fall apart, and does anything in this document support or contradict it?”

The first version gets you a generic list that could apply to almost any software company.

The second gets you a direct response to your specific reasoning, grounded in the document you’re actually analyzing.

You’re not asking Claude to think instead of you. You’re bringing it into the thinking you’re already doing.

Why most people never make this shift

It requires more work before you even start. You have to put your current view into words before it’s fully formed.

That feels awkward. Most people want answers before they’ve committed to a position.

But that awkwardness is where most of the value lives.

When you have to explain your thesis to Claude, you start noticing the gaps yourself. Half the time something clicks before Claude even responds. You realize mid-sentence that you’re assuming something you have no actual evidence for.

That’s not Claude doing the work. That’s you thinking more clearly because you had to articulate your reasoning out loud, even if only in a text box.

Claude then adds another layer. It engages with what you’ve written, picks at the weaker parts and occasionally brings up something you hadn’t considered. Not perfectly, not always. But consistently, on any company, at any hour, without trying to protect anyone’s feelings or defend a prior call.

That’s what a useful thinking partner does.

The prompt I actually use

Here’s the structure I run whenever I’m building a view on a company:

Read more