Hi everyone,
When you’re making things like carousels, decks or emails with AI, the words change every week but the layout and branding don’t.
Yet most skills ask the model to rebuild all of it from scratch every time, which wastes tokens and leaves room for mistakes.
We found plenty of that last Thursday, when the team and I spent a training session cleaning up our own Claude skills. The fix was simpler than we expected.
Let a script do the boring part

Now we split the work. A script handles the fixed part, a template handles the look, and the AI fills in the words and ideas, which is the part it's actually good at.
We saw three wins straight away. The result comes out the same every time, it uses fewer tokens so there's more usage left for real work, and the brand stays consistent.
Kubi made that last point in the session. Ask an AI to put the logo top left and sometimes it moves it, resizes it, or invents a whole new logo. A template puts it in the same place every time.
Kubi also answered the obvious worry, which is that templates get boring. Give the skill three or four templates and let it rotate: "you used number 2 last time, try 3."
We deleted one of our skills
Our daily blog automation was turning out mediocre drafts, and it sometimes scored them only 6 out of 10. They took so much editing you had to wonder why we bothered.
We'd built a weekly skill to tidy up the content calendar and archive the weak ones.
It worked, but as I said to the team, if you put a band-aid on the real problem, you’ll end up needing stitches anyway. So we’re going back to the automation itself and fixing the instructions to make sure it writes drafts worth keeping.
If you ever catch yourself building a skill to clean up after another skill, that's a sign that one should be fixed or deleted.
We've also started moving our team skills into Claude's organization library. That means team members can share and update a skill without going near GitHub. When one of us builds a skill that works, the rest of us use it instead of rebuilding it.
From this fortnight's trainings
The same advice came up in almost every client session over the last two weeks: plan for 7 to 10 rounds before an agent or a skill is reliable. Use it two or three times, write down where it slipped, and fix the instructions. Nobody gets it right on the first try, and that includes us.
The other question I heard in nearly every Copilot cohort was "should this be a notebook or an agent?" Here's the rule I give. A notebook is a project, with a topic, a start and end date, shared files and colleagues. An agent is a job description: one task that repeats, with instructions you can share. If it's a one-off task, you need neither.

A few smaller tips that people liked:
Put square brackets in a saved prompt, like [client name] or [audience]. People and AI both read them as blanks to fill in.
Paste an old chat that went well and ask Copilot to turn it into a reusable prompt.
Build the whole outline in chat first, then make the deck. Fixing a deck slide by slide is the slow way round.
Worth knowing this week
Claude Fable 5.2 is out. Anthropic released Fable 5.2 on November 9. It scores above Opus 5.5 on most of Anthropic's benchmarks and costs $8 per million input tokens and $40 per million output tokens on the API. As a launch offer, Pro and standard Team seats can use it for up to half of their weekly limit. Anthropic still recommends Opus 5.5 for everyday work.
I've been testing it this week from a fresh angle. I point it at things I've already built and ask what it would change, and it has already found improvements in my systems, my skills and my local setup.
Gemini 4 Carbon is out. Google released it on November 10 for paying developers and enterprises, and it's rolling out to Google AI Ultra subscribers in the Gemini app. Until November 30 it costs $2 per million input tokens and $10 per million output tokens, and after that $4 and $20. There's also a preview of /swarm in Antigravity, where one prompt starts a team of agents you can watch at work.
Copilot this month. Microsoft's November 4 release notes add an agent that builds SharePoint lists from a prompt, a usage report for agents that's coming to the admin center, and AI Workflows in the Teams Workflows app. AI Workflows are scheduled Copilot prompts. In a session this week I also saw skills you can call in normal chat with / or @. That isn't in Microsoft's notes yet, so it may not be in your tenant.
Anthropic is going public. Anthropic started its IPO roadshow on November 10, with a proposed Nasdaq listing under the ticker ANT.
Wed December 2: Copilot 101
Copilot 101 is three weeks away. It's a 60-minute live session for CEOs, COOs, CFOs and private equity teams who already pay for Copilot. I'll show the setup we teach, how to pick the first real task, and how to check an answer before you trust it.
It's live on Wednesday, December 2, at 9:30 AM ET, 2:30 PM GMT and 5:30 PM in Turkey. When you register, you can tell us one task you'd like AI to do, and I'll pick real tasks from those for the live demo.
One thing to try this week
If you’ve built a few skills of your own, here’s a quick way to find out which parts should be a script or template instead.
Claude has a skill called Skill Creator, and it's the best teacher I know for this. You could open it and ask something like this:

There's no need to change anything straight away. Read what it suggests, pick the one idea that would save you the most rework, and try just that one.
Q1 individual cohorts
The Q1 waiting list is still open. In Q1 you'll be able to join one of our cohorts on your own, without your company signing up. Reply "Q1" and I'll add you.
From the blog
Who should read it: operating partners who are writing AI into a 100-day plan.
TLDR: A three-phase plan, from naming an internal AI lead to a day-100 report the board can act on.
Which of your skills or prompts gives you a slightly different result every time?
Hit reply and tell me. I read every reply.
Thanks for reading.
Tim
