This summer at Shark Skin we brought on four very sharp interns. Not uncommon for a venture organization. Then we did something less common: we hired each intern their own intern.
It wasn’t a selfless act. I wanted the interns to be productive, and I wanted them to learn the most valuable skill of the next decade — delegation.
To be clear, the interns’ interns were not the coffee-fetching, badge-wearing, needs-two-weeks-to-find-the-bathroom variety. Their interns had read everything. Every book, every research paper, every codebase, every business strategy ever written down. They speak every programming language. They can write, analyze, synthesize, design, argue, and build. They don’t sleep, they don’t get distracted, and they have zero ego about being wrong.
Each intern’s intern was named Claude. We put every one of them on the Max plan — about $200 a month, more than most firms spend on an intern’s entire software stack. It was the best money we spent all summer. (besides the summer party of course)
And this wasn’t AI summer camp. The interns worked inside a lean-startup process — daily standups, weekly sprints, real deadlines — across two live companies: The Everlasting, a platform that helps families capture and preserve life stories, and a second company we’re building in personal cyber security company in stealth mode — corporate-grade defense for high-net-worth families. Real products. Real customers. Real consequences when something shipped broken.
What they shipped
Here’s a sampling of what the four interns and their four interns have shipped in the last six weeks:
A production payment platform. Stripe-integrated, connected to our internal systems through MCP — the live connection that lets Claude read and act on real company data. Processing real transactions today. Built entirely by the interns.
An account management system. A separate build, also MCP-connected, live in production, letting users manage their own accounts. Also entirely intern-built.
A documentary production pipeline. A full workflow for editing and producing long-form documentary films with AI doing 80%+ of the work — built on Shotstack, Google AI Studio, ElevenLabs, and Tavus, all orchestrated through our own MCP. This was a major design and execution task.
A fully automated content agent. Every morning it sweeps the industry, selects the stories that matter, and drafts original commentary with links back to every source. The content is publishing-ready, no human in the loop. It is one of the most productive agents I have seen anywhere.
A scam-detection agent. A consumer facing, drop in any suspicious text, email, or voicemail, and it diagnoses the likelihood you’re being scammed, in seconds.
A working CRM, built in partnership: multiple drip campaigns, client journeys, and a feedback loop that collects data to improve those journeys and lift engagement rates.
An iPhone app prototype, designed and built from the ground up in a matter of weeks.
A backend admin system where all of our documentation and product roadmap lives, tied directly to our MCP. I’ve written before about MCP as the plumbing of the AI-native company. The institutional knowledge doesn’t live in someone’s head. It lives in the nervous system of the business.
None of this happened because the tools are magic. Here’s the catch: Claude is only as useful as the direction you give it. The smartest intern in the world sitting in a room waiting for vague instructions is still just a person sitting in a room. The leverage only materializes when you learn how to direct it — specifically, clearly, and with enough context that it can actually work at the level it’s capable of.
So the real adventure this summer wasn’t the projects. It was learning to delegate. Direct Claude well, and your output becomes unrecognizable compared to what you’d produce alone. Learn to build effective agents, and your reach extends far beyond the normal grasp of an internship. Direct it poorly, and you’ll spend all summer talking to the most capable intern on earth and getting generic, meaningless work that will convince your boss you are lazy and incompetent.
This connects to my emerging theory of employment: the number one prerequisite for any job now — nothing else is even close — is intellectual curiosity.
What I’ve learned over the last nine months is that for anyone genuinely interested in learning and building, it has never been easier.
Over the next three days I’m publishing the ten rules we gave our interns the day we introduced them to their interns, in three parts: how to talk to Claude, how to think with Claude, and which tool to reach for when. Parts I and II apply to anyone doing anything with Claude — from an accountant to a stay-at-home dad running the household. Part III is where the engineering lives.
Part I: How to Talk to Claude
1. Context Is Everything
Key takeaway: Garbage in, garbage out. The more specific you are, the better Claude gets.
Before you ask Claude to solve anything, ask yourself: “What would a smart person who’d never heard of my company need to know to do this well?” That’s roughly what you need to give Claude.
Claude works by pattern-matching on what you provide. If you ask it to “make this better,” you’ve given it nothing to work with. It’s like asking someone to fix your car without telling them what’s wrong.
What to include in every prompt:
The specific context Claude needs (what’s the goal? who’s the audience?)
Your constraints (tone, length, format, any rules)
What you’re optimizing for (clarity? persuasion? brevity?)
Reference materials (existing copy, competitor examples, brand guidelines)
Bad: “Write copy for our homepage.”
Better: “Write a 150-word hero section for theeverlasting.ai targeting adult children 40–65 with aging parents. The emotional hook: most families wait until it’s too late. Tone: warm but elevated, no jargon. Short punchy sentences at emotional peaks. Here’s our current homepage for reference. Here are three competitor examples of what NOT to do.”
The second one isn’t longer because you’re being verbose. It’s longer because you’re being useful.
2. Ask What Good Looks Like — and What Failure Looks Like — Before Asking for Output
Key takeaway: Before you ask Claude to help you build something, ask Claude what actually works for that thing — and what commonly kills it. Then apply both to your specific case.
Most people skip this step. They jump straight to “help me build X.” What they should do: understand the craft first, then apply it.
This does two things. It grounds your work in reality, not just your narrow situation. And it produces better downstream output, because Claude is now anchored in what “good” looks like in this domain — and is building against real failure modes instead of guessing.
The move is three prompts, not one:
“What are the best practices for: email re-engagement campaigns / user research interviews / landing page design? What actually works?”
“What are the three most common mistakes people make with this? What separates good from mediocre?”
“Given those best practices and those pitfalls, here’s our specific situation. How should we approach it?”
Example: Don’t open with “Help me design interview questions for beta users.” Open with “What makes a user research interview actually work, and what ruins one?” Then: “Given that, draft 8 questions to understand why users drop off after session 3.”
Our interns made this their default first move. It costs two extra prompts and saves you from optimizing for the wrong thing.
3. Constraints Beat Open-Endedness
Key takeaway: Give Claude guardrails. Constraints force better choices.
When you leave Claude completely open, you get outputs that are generic, safe, and forgettable. When you constrain the problem, Claude has to think harder.
Types of constraints that matter:
Length (150 words, not “concise”)
Audience specificity (founders in the AI space, not “tech people”)
Tone (warm but elevated, contrarian, direct)
Format (three bullet points, a narrative arc, a numbered list)
What NOT to do (here are three bad examples of this)
Bad: “Write a landing page that converts.”
Better: “Write a landing page hero for a specific audience that emphasizes one emotional truth without sounding like every other legacy service out there.”
The constraint isn’t busywork. It’s the difference between Claude generating a template and Claude generating something that actually fits your business.
Tomorrow, Part II: what to do when Claude’s first answer misses — which it will — and how to stop it from brilliantly solving the wrong problem.
And we’re already building next summer’s class. This is not a resume-padding internship. If you know a high school junior/senior or college student — 18 to 21 — with real intellectual curiosity, applications are open: [APPLICATION LINK]