Friday Field Notes.

What to do

If you run a Shopify store, pick the one product page you've always known was your weakest. Open Cowork, pull in Claude Design, rewrite the hero copy, regenerate the image, update the metadata, push it live in one pass. That's the new loop. Then do it for the next ten pages.

If you run a Shopify partner agency, the pitch is services-as-software layered on AI agent architecture. Your Shopify expertise is the part an agent can't replicate. The sharpest offer you can bring a client is "we'll design the agent system that runs your store."

What happened

On Wednesday, Anthropic shipped Opus 4.7 (announcement). Real capability jump on the things that have been holding agents back: long-running tasks, following multi-step instructions precisely, verifying its own output before reporting back.

Two days earlier, Garry Tan published "Thin Harness, Fat Skills" (read it). His central claim: the people getting 2x productivity out of AI and the people getting 100x are using the same models. The difference isn't intelligence. It's architecture.

Why it matters

Opus 4.7 is a real capability jump. Use cases that were marginal a week ago — an onboarding agent that drifts at step seven, a research loop that summarizes but doesn't verify, a client intake that handles the easy 80% — are now in range. That's not hype. That's a model that stays on task long enough to finish.

But the jump only shows up if you've already built the system. A system isn't a clever prompt. It's a set of reusable skills — markdown files that teach the model how to handle a specific kind of work — with reliable plumbing underneath. You write the skill once. It runs tomorrow, next quarter, next year. Every time the model underneath gets smarter, every skill you've written gets smarter with it.

If you spent last month on that, Opus 4.7 just lifted all of it without you touching a line. If you spent it on one-off prompts, those still work — they just don't stack. You're getting answers, one at a time, and none of them become an asset.

Every model upgrade is a free upgrade to the system you already built.

What to do

The question worth asking isn't which model. It's this: am I building a system that will compound over time and upgrade with every model release automatically? That's the question that separates a tool you rent from a compounding asset you own.

Client signs. Then what?

If the answer depends on who's available, what they remember, and which folder the template is in — that's the problem. Not the tools. The process lives in your head and it runs differently every time.

An onboarding agent encodes your actual process into a system. Client intake comes in, the agent produces your kickoff materials, schedules your follow-ups, and writes everything in your voice — because it knows your business, not because it's filling in a template.

Why it works

The agent reads from a context file that knows your business. The welcome packet sounds like you. The follow-up references details from the actual intake, not "hope things are going well." The internal handoff doc hits the same format every time.

Why it compounds

After five clients, you start noticing patterns — which questions always come up in week one, which deliverables clients actually open, where things stall. You update the skill file. The tenth onboarding is better than the first. That's a system, not a template.