One of the more useful shifts in AI has nothing to do with smarter models.

It’s getting AI to remember how you want a recurring job done, so you don’t have to explain it again next Monday.

Google made that idea much more explicit this week.

Google started rolling out Skills in Gemini, which let you save instructions once and reuse them whenever the same kind of work comes back.

Skills can include reference files, work together, and even trigger automatically when Gemini recognizes a matching request. Google says Skills will begin replacing Gems for personal accounts in November.

This is more interesting than another prompt feature.

Think about the useful prompts sitting in your notes, bookmarks or old chats. If one consistently produces good work, the next step isn’t writing a slightly better prompt.

It’s making the instructions persistent.

A sales-call analyzer. A proposal reviewer. A weekly reporting process. A research brief. A client-onboarding checklist.

The useful asset is increasingly the repeatable operating logic around the model, not the clever sentence you typed into it.

My rule: if you’ve given AI substantially the same instructions twice, ask whether they should become reusable.

Shopify quietly moved AI shopping another step forward this week.

Browser-based AI agents can now use Shopify’s checkout tools to read and update checkout details, then submit an order after the buyer confirms it. Combined with its existing storefront and cart tools, an agent can assist from product discovery through order confirmation.

Shopify says merchants don’t need to configure anything new for the checkout capability.

I wouldn’t spend your weekend optimizing a store for robot shoppers.

But I would notice the direction.

For years, ecommerce has been designed around a human opening pages, comparing products and clicking buttons.

Increasingly, some of that buying process may be delegated to software.

That makes boring ecommerce housekeeping more valuable: descriptive product titles, complete variant data, obvious pricing, current availability, sensible policies and copy that clearly says who something is for.

That’s increasingly a form of machine-readable merchandising.

If an agent can’t confidently determine what you sell, who it’s for and what it costs, that eventually stops being just a website problem.

It becomes a distribution problem.

Anthropic published research this week estimating that existing robot capabilities cover 74% of physical work tasks in the U.S. in at least some setting.

Important caveat: that’s an exposure estimate based on what current robots appear capable of doing, not a field test showing robots successfully performing three-quarters of physical work in real businesses.

Then comes the useful number.

The researchers estimate robots are currently cost-competitive with human labor for only 0.3% of work. Many also need highly structured environments that real workplaces don’t provide.

That gap is a useful filter for almost every AI demo you see.

Capability is not the same thing as a business case.

When evaluating automation, don’t stop at:

Can AI do this?

Measure what happens after software cost, implementation, supervision, mistakes and exceptions.

A less impressive automation that reliably removes 30 minutes of paid work every day may be much more valuable than an autonomous-agent demo that technically does everything.

Here’s one way to turn a recurring prompt into something you only have to set up once.

Create a saved AI setup called Meeting Follow-Up that turns a meeting transcript into:

You can build it in ChatGPT Projects, Claude Projects or Gemini Skills/Gems by saving a set of instructions that tells the AI exactly how you want this job done.

Then, after each meeting, open your Meeting Follow-Up, paste or upload the transcript and type:

“Follow up on this meeting.”

That’s the whole recurring command.

Because the detailed instructions are already saved, you don’t have to explain the job again every time.

Copy the Meeting Follow-Up setup →

Includes the exact instructions plus simple setup steps for ChatGPT, Claude and Gemini.

If you normally spend 10–15 minutes turning meeting notes into follow-up, four meetings a week makes this a reasonable candidate for about an hour of admin to compress.

If enough people want it, I’ll make it.

Know someone who spends too much time cleaning up after meetings? Forward this to them.

Talk soon,
Jason