This is Part II of The Intern’s Intern. The premise, if you missed it: this summer we gave each of our four interns their own intern — Claude, on the $200-a-month Max plan and in six weeks they shipped production systems for two live companies. Part I covered how to talk to Claude: context, craft, constraints. This one covers how to think with it.
Part II: How to Think with Claude
4. Name Your Assumptions. Then Go Get the Data.
Key takeaway: The biggest mistake people make with Claude has nothing to do with prompting. It’s skipping the scientific method — solving before validating. Define the key assumptions underneath your problem, then use Claude to test them with data. Only then let it optimize.
For nearly two decades at Dreamit —we ran hundreds of startups through the same discipline: before you build anything, name the key assumptions your business depends on. (Steve Barsh is the master of this) Then go get data on each one. The startups that succeeded weren’t the ones with the best ideas — they were the ones that found out fastest which of their assumptions were wrong.
We didn’t invent that. It’s the scientific method: hypothesis, evidence, revision. The same loop that took humanity from caves to air conditioning in a couple thousand years. All we did at Dreamit was refuse to let founders skip it.
Claude is the best assistant the scientific method has ever had. It can surface the assumptions hiding inside your plan, design the cheapest test for each one, and analyze the results without fatigue and without ego. But it will only run the loop if you make it. Left alone, Claude will brilliantly solve whatever problem you hand it — including the wrong one.
So before asking “How do we get users to complete more videos?” ask: “Here’s our plan, and here’s our hypothesis for why users drop off. List every assumption we’re making. Which one, if wrong, kills us? What evidence would prove or disprove it?”
Here’s what the same discipline looks like on a live metric:
Weak: “This email isn’t converting well. Fix it.”
Strong: “We’re sending this email [paste it]. Conversion is 6% against a 15% industry benchmark. Our hypotheses for why, ranked in order of suspicion: (1) the offer is buried below the fold, (2) the body copy speaks to the wrong pain point, (3) we’re sending to too broad a segment. We’ve already A/B tested subject lines — that one’s ruled out. What hypotheses are we missing? Which is most likely given the email in front of you? And what’s the cheapest test to prove or disprove each one — before we rewrite anything?”
Notice what the strong prompt does NOT do: it doesn’t ask Claude to fix the email. It states the hypotheses ranked by suspicion, shows the evidence already gathered, asks Claude to attack the list, and demands a test plan. The rewrite comes after the data — that’s the whole discipline. The weak prompt buys you a guess. The strong one buys you an experiment.
Claude can reason about a specific case with named hypotheses. It can’t reason about abstractions or read your mind about which assumption you’re betting on. The two failure modes are mirror images: solving the wrong problem precisely, or solving the right problem vaguely. The hypothesis-first habit kills both.
5. Iteration Is the Skill
Key takeaway: One prompt rarely lands. Learn to read the miss, describe it specifically, and iterate. The back-and-forth is where the magic happens.
Most people ask Claude once and take whatever comes back. That’s leaving 90% of the value on the table.
Claude can’t read your mind. If the output misses, your job is to describe the specific miss, not just say “try again.”
The bad iteration loop:
You: “This isn’t right.”
Claude: “How should I change it?”
You: “Make it better.”
Repeat forever.
The good iteration loop:
You: “This reads like a generic productivity app, not a legacy service. The emotional stakes aren’t there. Rewrite to emphasize why capturing now matters — the fear of losing someone before you truly knew them.”
Claude: [rewrite]
You: “Better. But now it’s too dark. Keep the emotional truth but shift the ending to hope — what they gain, not just what they lose.”
Claude: [rewrite]
Notice the difference? The second loop actually tells Claude what was wrong and what to fix. It’s a conversation, not a guessing game.
6. Use the Steel Man Technique
Key takeaway: Before Claude critiques your idea, ask it to make the strongest possible case FOR your idea first. Then ask for the strongest case against.
You’ll get much more useful pushback, because Claude has already committed to understanding your thinking, not just punching holes in it.
This takes an extra prompt. It’s worth it.
The final part tomorrow: which tool to reach for when — Chat, Code, agents, skills, design — and where the gold actually gets made.
Meanwhile: we’re building next summer’s intern class. Highly selective, and one attribute matters above all — intellectual curiosity. Ages 18 to 21. Applications are open: [APPLICATION LINK]