Twenty-eight skill files, one router to pick between them, and not a single original idea in the whole stack. Which means the one thing you cannot hand to a machine is the exact thing you get paid for.
The Briefing
1. Zillow moved into the chat window. On October 19 Zillow switched on a connector for Claude. Your buyer can now search listings, check a home value, book a tour and ask about financing without ever loading Zillow's site. Property is just the category that went first...
Your move: Write down the three questions a buyer asks before they would ever visit your site. Those answers need to live somewhere a machine can read them.
2. The intake data says humans still book the next step. Smith.ai went through 3.2 million law firm calls. Qualified leads booked a next step 59% of the time with a human, 52% hybrid, 33% with AI only. Their calls, their report, so weigh it as a vendor number.
Your move: Find where AI sits in your intake. If it is inside the qualifying conversation, move it to the hour before.
3. Some of your traffic is not a person. DataDome's Q3 report counts nearly 25 billion AI agent requests. Meta's Muse browser sent 29.1 million of them without identifying itself as an agent at all.
Your move: Open last month's analytics before you judge another conversion rate... check direct and "other" first.
4. Nearly half of US buyers dodge AI in ads on purpose. Auros surveyed 4,000 adults across the US, UK and Australia. 44% of the US respondents said they go out of their way to avoid brands that use AI in advertising. Research by Talker Research, so treat it as directional.
Your move: Keep AI in the back room. Research, drafts, qualifying. Keep your name and your face on what ships.
The Build
Twenty-eight skill files. One router that picks between them.
Zero real ideas of its own.
I spent this week deeper inside my own stack than I have been all year. The file at the centre of it is SKILL-ROUTER.md, which does one job: read the request, pick the right skill, hand the work to it.
Sales copy goes here.
Offer stack goes there.
Market research, that one.
Webinar script, the other one.
It picks well. That is the part worth saying out loud, because the routing itself is genuinely good...
Then I noticed every row in that table starts with the same four words. When the user asks.
Twenty-eight arms. All of them waiting for a real signal…
No row for what is worth asking for.
No row for the angle.
No row for knowing the market is bored of the thing that worked last year.
The deeper I go with this stuff, the clearer it gets. It builds.
It doesn’t think.
It is a doer, and what it repeats is old, or poor thinking, delivered faster than you can type.
But the worry everyone keeps airing is the wrong worry. The line does not run between you and the machine. It runs between thinking and clicking.
So be honest for a second... what did you actually think about last week?
AS A MARKETER I'M SAFE.
AS A CLICKER I'M COOKED.
So I stopped asking the stack to do the job and started making it show me which half of the job was mine…and what was “it’s”.
Here is the job I am about to hand you: [DESCRIBE THE JOB]
Before you do any of it, split it into two columns.
ARMS: the parts that are execution. Known inputs, known format,
a right answer I could check.
BRAIN: the parts that need a decision only I can make. The angle,
the market, the tradeoff, the thing to stop doing.
Rule: if a step needs taste, a bet, or knowledge of something that
is not written down anywhere, it goes in BRAIN. Do not be generous
to yourself.
Now do the ARMS column. Leave BRAIN blank and hand it back to me
as questions.
Then tell me which item you put in the wrong column, and why you
were tempted.Where it broke: It is generous to itself. Ask a model to sort its own work and it will file judgement calls under execution, because filing them there lets it keep going.
That is the exact move that produces work which reads competent and lands dead. I have no numbers on this yet, so I am not claiming the two columns lift anything until a dated result says so…but it will come.
What held: The router. Twenty-eight arms, each one better at its job than I am at that job, all of them reachable in one line. And the split as a standing habit, run before the work instead of after it.
The Octopus. Eight arms, one brain. The arms are brilliant, they work independently, and not one of them has ever decided where to swim.
The Play
If you are a marketer who has been quietly worried for a year now, this is the hour that settles it. The ones who have been doing this a long time already know how it ends. The job is not under threat. The performance of the job is.
Open last week's calendar and your sent folder. Then run this:
Here is everything I worked on last week. [PASTE THE LIST]
Sort every item into ARMS or BRAIN.
ARMS = execution. Known inputs, known format, a right answer that
can be checked by someone who was not in the room.
BRAIN = a decision only I could make. The angle, the market, the
price, the bet, the thing to kill.
Then give me three things:
1. The hours in each column.
2. The three ARMS items I should never touch by hand again, and
which tool or skill takes each one.
3. The one BRAIN item I rushed because the ARMS work ate the week.
No encouragement. If the BRAIN column is thin, say it plainly.
Then tell me which item you were unsure about, and what you would
need to know to place it properly.The rule people skip: do it on last week, not on a normal week. There is no normal week. Memory sorts in your favour and your calendar does not.
Last week this was something I couldn’t name. Today it’s two columns I can point at and count.
Try this today. Write ARMS or BRAIN beside every block in last week's calendar, then count the brain hours. That number is your actual job. Everything else is an arm waiting for a signal.
James
P.S. Marketers… AI is not coming for your job. It is coming for the pretenders. The ones who’ve been doing this twenty years worked that out in the first week AI came to life.
Growth Partner for founders stuck between $1M and $20M | I find the money you're sitting on and build the machine that collects it | $1B+ in client results.