What You’ll Learn

James has written thousands of Claude skills over the last two years. He isn’t going to write another one. In this session he explains why, and what he does instead: a five-part briefing that tells an AI employee what you want done, not how to do it.

If you have been collecting prompts and skills and wondering why they keep going stale, this is the shift that fixes it.

Why Skills Keep Going Stale

The trigger was an interview with Boris Cherny at Y Combinator’s Startup School. Cherny is the person behind Claude Code. His advice to a room of founders was blunt: delete your instruction files and start fresh.

His reasoning is the part worth sitting with. Every instruction you write today is scaffolding propping up today’s model. The model keeps improving. The scaffolding does not.

“We don’t build the model for today. We build the model for 6 months from now.”

James puts a house rule on it that Anthropic apparently keeps on the office wall: never bet against the model.

💡 In Plain English: A skill is a very detailed set of instructions for doing a job a specific way. When the model gets smarter, your detailed instructions stop being help and start being a limit.

The Shift: From Procedure to Job Description

The old way was to spell out every step. Do this, then this, then this. You wrote it once as a skill so the tool would never forget, and you could count on it later.

The new way treats the AI as a new hire rather than a machine. You don’t hand a capable new employee a script for every keystroke. You tell them the outcome, what good looks like, and what to check with you before doing.

James is explicit that this is not about a virtual assistant you micromanage. It’s about briefing someone who already knows how to work.

The Five-Part Handbook

This is the briefing James now gives every AI employee in place of a skill.

1. Here’s the job

State the outcome, not the steps. “I want my blog SEO optimised.” “I want a new product launched.” “I want content marketing done for the week and actually posted.” You are naming the finish line and letting the employee find the route.

2. Here’s what good looks like

This is the one most people skip, and it’s why the first attempt disappoints. If you don’t define done, you get a blog post with twelve words, no call to action, and no optimisation — technically delivered, actually useless. Give the criteria for acceptance, not the method for hitting them.

3. Here’s what you never do without asking

Guardrails. Draft the email but don’t send it. Write the post but don’t publish. Anything that spends money, goes to a customer, or can’t be undone stops and waits for you.

4. Here are your files, tools and history

Shared business memory. Not just the chat you had yesterday — your brand, your ideal customer, your products, your pricing, how you solve problems for the people you serve. This is the context that makes the output yours instead of generic.

5. Go

Then get out of the way and watch what happens.

✓ Check Your Work: Take one skill you rely on. Can you restate it as outcome, acceptance criteria, guardrails and tools — without describing a single step? If you can, you have a job description. If you can’t, you’re still writing a procedure.

Try It On Something You Already Own

James’s challenge is concrete. Pick an existing skill. Delete it. Ask for the same outcome using the five parts above. Then watch where the model gets stuck, how it approaches the problem, and whether its route was better than yours.

“They’ve got more training than we do on the best way to do things.”

He admits the habit he struggles with most is asking for too little. His correction: ask for ten times the result you normally would. It’s an AI employee with access to unlimited knowledge and real tools, not a person you need to protect from a heavy week.

Where the Manager Layer Comes In

A fair objection: won’t the model eventually just do all this itself? James’s answer is that a single agent handling everything becomes generic fast, because sales, marketing and delivery each carry different memory and different playbooks.

That’s the argument for departments with managers. A manager knows what day it is, what runs on a cadence, what happened last time it ran, and how that should change what happens next. That belongs to a department, not to one universal assistant.

⚠️ Worth knowing: This is a trade, not a free win. James is candid that it doesn’t work perfectly the first time. The thinner your brief, the more questions come back. Getting good at outcomes and acceptance criteria is the skill that replaces writing skills.

Key Takeaways

• Anything you write to prop up today’s model is temporary — the model improves faster than your prompts do.
• Replace step-by-step skills with a five-part handbook: the job, what good looks like, the guardrails, the tools and memory, then go.
• Defining “done” matters more than defining “how” — vague acceptance criteria are why first attempts disappoint.
• Keep human gates on anything that publishes, sends or spends; drop the choreography in between.
• Test the idea by deleting one skill you rely on and asking for the outcome instead.

Your Next Step

Pick one skill this week. Delete it. Write the five-part brief instead. See where it gets stuck — that’s your real feedback, and it’s worth more than another prompt.

The Campus AI OS is free to download and gives you Dean plus a starting roster of AI employees to practise this on: trainingsites.io/os