Your AI is a new hire. Treat it like one.
By Nick Basile
The harness is the point
Imagine you just hired someone. You're excited. They're smart. But on day one, you don't just hand them a laptop and walk away.
You sit them down. You tell them who your customers are, how you talk to them, and what the non-negotiables are. You give them logins to your calendar, your files, your email. You set up a rhythm -- a Monday standup, a Friday wrap, an end-of-day check-in.
All of that stuff -- the context, the access, the rules, the rhythm -- is the difference between a new hire who's useful in their first week and one who flounders for a month.
Here's what took me a while to realize: AI is exactly the same.
When you open ChatGPT or Claude and type a question, you're talking to someone who just walked in off the street. They're smart. They're fast. They also have no idea who your customers are, what your business does, or how you'd want them to handle a refund email. Of course they give you generic answers. You haven't onboarded them.
The stuff you'd set up for a new employee has a name in the AI world now. It's called a harness.
A harness is every piece of context, access, and rules you wrap around the AI so it can actually do useful work in your business. It's the onboarding packet. It's the company handbook. It's the login credentials and the house style guide. Without a harness, the AI is a smart stranger. With a harness, it's a coworker who knows where things are.
And this week, three things happened that made it clear this is where everything is moving.
The CEO of Box (one of the big enterprise file-sharing companies) said out loud that enterprises are going to start dropping vendors whose products aren't easy for AI to plug into. An engineering lab called LangChain published a framework with a line I can't stop thinking about: Agent = Model + Harness. And Anthropic (the company behind Claude) quietly pushed three products out of testing and into full release -- a new collaboration app, a Microsoft Word plug-in, and a batch of upgrades to their developer tool.
Read those three together and the message is hard to miss. The AI itself isn't where the value lives. The setup around the AI is.
Here's the good news for small business owners: you do not need to buy another AI product to take advantage of this.
You need to onboard the AI you already have. That's what we've been doing on Learn AI with Nick from day one; I just didn't have the word for it yet:
Start with a simple prompt.
Turn the prompt into a reusable skill.
Chain skills into an automated workflow.
Step back and realize you've built a harness.
Every rung on that ladder is the same move. Give the AI more context. Give it more access. Tighten the rules. Keep going until it fits the shape of your business instead of the other way around.
So what does a harness actually look like for a one-person shop? Something like this:
A
CLAUDE.mdfile (just a plain text document) that tells the AI who your customers are, what tone to use, and what's off-limits.A few skills (reusable instructions) that encode your standard operating procedures. How you write a proposal. How you respond to a refund request. How you prep for a client call.
A scheduled check-in that reads your inbox every morning and drafts replies you can approve in five minutes with coffee in hand.
Permissions that say yes to drafting, no to sending.
That's it. That's the harness. No magic, no genius. It's the boring, specific, unglamorous work of telling the AI what you know and what you want.
So here's the shift for the rest of the year. Stop asking which AI tool to buy; start asking what your team already does and how to wrap the AI around it. The vendors that matter going forward are going to be the ones that make that wrapping easy. The businesses that compound are going to be the ones that actually do it.
This week's links
Our Story — Mindstream
Matt Village and Adam Biddlecombe have shipped a newsletter every single day since June 2023. Every. Single. Day. HubSpot acquired them 17 months in, and they're now at 230,000 subscribers with a three-person editorial team. Here's the piece worth stealing from their own origin story: their welcome sequence is written in three voices -- Matt, Adam, and their first hire Maria -- across multiple emails. By the time it ends, you know the crew. You're no longer subscribed to a publication; you're subscribed to people.
Design Buddies — Grace Ling
No VC. 150,000 designers in the community. And one rule did most of the heavy lifting: before you can ask for feedback, you have to give feedback to three other people first. That single rule turned a place people showed up to consume into a place they showed up to reciprocate. If you run a community of any size -- a Slack, a Discord, even a client channel -- ask what one rule of yours could create that kind of gravity.
Claude Code Changelog — Anthropic
The official changelog is the best place to see what Anthropic is shipping, so you can decide what to pull into your own setup. Standouts from the last few weeks: /ultraplan for sketching a plan in the browser and running it in the terminal, a Monitor tool that watches long-running jobs without burning tokens, and a background-monitor manifest so plugins can auto-arm on session start. If you live in Claude Code, skim this weekly.
The Anatomy of an Agent Harness — LangChain
Vivek Trivedy wrote the clearest piece I've read on the concept I opened with. Agent = Model + Harness. If you want to understand why context files and skills and schedules matter more than which model you picked, this is your starting gun. Technical but readable in one sitting.
Supabase published a set of Agent Skills -- open-source instructions that teach any AI coding agent how to build on Supabase correctly, covering Row-Level Security, schema, and the CLI/MCP workflow. This is exactly what Box's CEO was talking about: a vendor making itself easy for agents to use. It's also a great template if you want to see what an "agent-friendly product" actually looks like under the hood.
Glenwood — Justin Barber
Justin Barber was the first designer and then Head of Design at Applied Intuition. In February 2026 he quit to start Glenwood, a creative studio he calls the feelings company of Palo Alto. His line is the one I keep thinking about: "margin can't be my primary criteria for making product decisions." Read when you need a reminder that not every business has to chase scale.