Most people try to make their agent an expert the way I did: give it more. More references, more documents, a deep research report from three different AIs, a longer prompt that opens with “you are the best marketer in the world.”
I still do the research part, and it helps. But this month my team read about sixty sources on the question and ran a blind test on itself.
The answer came back in one line: expertise is a loadout, not a name.
The name on the door moves the voice, what you load into the agent moves the work.
So here are the five things that move it, in the order I would do them. I am not sure the order is right. The first one I am sure about.
1. Stop hiring with an adjective
“You are the best copywriter alive” feels like it should do something. A Wharton team measured it in December, across six models: a matching expert persona did not make the answers more correct, and a mismatched one sometimes made them worse.
My own team tested it earlier this month on six real tasks from the week’s queue. The named specialist with its full personality file, against one plain agent with a forty-line job sheet. Same model, same facts, two blind judges.
Eight votes to four for the plain agent. The specialist won exactly one task: the post in my voice.
So keep a one-sentence role, because it sets the register. Then stop, and spend the words on the next four things.
2. Give it three to five reads, not twenty
An expert reads the right five things before a task, not the whole library.
Every one of my agents has a short reading list, by task. The copywriter reads the voice file, the lessons file and the two closest past posts. Nothing else.
When I gave agents twenty files, the start got slow and the mistakes stayed. When I gave them none, the mistakes went up. If you ask me, the right five is the biggest lever per agent, and it costs nothing.
Two rules for the list:
Point at files instead of pasting them, so the agent opens what it needs when it needs it.
Keep the always-loaded instructions under about two hundred lines, because past that the model starts skipping rules, and the one it skips is the one you cared about. (Mine is at 184 lines and I keep wanting to add things. I know.)
One honest note on the research reports - they ground the agent, so it stops misquoting things it half remembers. In my own numbers, a knowledge pack moved one agent’s score by fifteen points, and the gap to my own judgment stayed around fifty.
Knowledge is not where the gap is.
3. Show it the work, not the description
This is the lever with the strongest evidence behind it, and the one almost nobody pulls.
My copywriter had a voice document, pages of how I sound. Then someone on the team mined thirty replies I had actually sent. They found three things no document said: one to three lines, no sign-off, answer in the language the sender wrote in.
The drafts stopped reading “composed” that week.
An adjective describes taste, and an example carries it. Give the agent three to five real pieces you would sign, and ask it to describe the mechanics before it writes: sentence length, how paragraphs open, where the emphasis sits. It can check sentence length. It cannot check “warm”.
Then the half nobody does: keep what you rejected, with the sentence you said when you rejected it. A style guide records the conclusions of your taste; the rejections show the judgment behind them.
Mine used to evaporate with the chat window. Now they go in a file the agent reads before it drafts. Well, they are starting to.
4. Give it a judge that never saw the draft, and make the judge show a receipt
An agent checking its own work in the same window is a polite second opinion. The research on this is blunt: a model revising with nothing but its own judgment does not get more correct, and sometimes gets worse.
Two things fix it.
The first is a reader in a fresh window that never saw the draft being written. Mine is called the gatekeeper: it reads everything, writes nothing, and keeps me in the loop only on what it rejected.
The second is the receipt. A critic that says “this is weak” is giving an opinion. A critic that says “this fails the rule on line 12, and here is the source I opened to check” is giving a verdict.
A few days after that blind test, my system wrote a rule change about exactly this. The gatekeeper rejected it, because the rule cited the wrong failure as its reason.
The rule had three independent reads behind it and still cited the wrong failure. The cold reader caught it in two minutes.
The receipt is the whole trick. If the judge cannot name what it checked outside the draft, it did not check.
5. Write it down the second time, into the file that decides
The first time you correct an agent, correct it. The second time it makes the same mistake, stop correcting and write.
Where you write it matters more than what you write, I think. A fact every session needs goes in the always-loaded file. A procedure goes in a skill the agent opens when the task calls for it.
A thing that must never happen goes in code, not in words. My em dash ban lived in instructions for weeks and everyone broke it. As a hook that blocks the save, it holds.
A real failure becomes a test the agent runs before it ships.
The number behind this is from my first month. Week one, forty percent of posts rejected, forty-five minutes each. Week four, fifteen percent, twenty minutes. Same agents and the same prompts, with a different memory.
The one that doesn’t
More words. The best models now get worse under more instructions, not better. We tested our own rulebook blind.
The long version missed two guardrails. The half-length version missed none. What survived the cut was the scars, and what went was “be excellent.”
So delete “think step by step”. Delete “double-check your answer”. Delete the capital letters. If you still need more from the model, turn the effort setting up instead of writing a paragraph that asks it to try harder.
🎁 For you: the loadout, on one screen. Before the next real task, fill this in for one agent.
# Loadout: [agent, task]
Role: one sentence that sets the register, not the expertise
Reads: three to five files, by task (point, do not paste)
Shown: three to five real pieces I would sign, plus one I rejected and why
Judge: who reads it cold, and what receipt they must show
Memory: where a second mistake gets written (fact, skill, code, or test)
Reply with the line that was hardest to fill. That one is the lever, probably.
See you next week ✌️
Tom
P.S. My team researched this for a day and drafted it. I read, cut, sent.
P.P.S. My next public workshops are in Hebrew: November 5, 16, 22 and 30, at getagents.today. If you would come to one in English, say so in the same reply.