You’ve probably typed “make it better” into a chat with an agent.
The agent does what you asked, and the draft comes back longer, smoother and more confident, and still not right. So you type it again, and what you keep getting is polished average.
There are 5 questions further down that fix most of this, and you can run them today. First, why “make it better” doesn’t work.
Why agents give you the middle
An agent has read more than anyone alive, and that’s the problem.
Ask it for a headline, a logo or a plan with nothing else to go on, and you tend to get the middle of everything it has read. That’s a lot of what people call slop. It isn’t broken, it just belongs to nobody.
It also has nothing at stake. It won’t stand in front of your client or read the angry reply, and it doesn’t carry last month’s launch that flopped, unless somebody wrote the lesson down where it reads.
“Make it better” asks the one thing it can’t know: better for whom, and by whose standard.
What taste is
Taste is a no with a reason.
Anyone can say “I don’t like it.” Taste sounds like “this opens with a question, and our readers skip questions,” or “this is right, but it’s right for somebody else.”
It has two halves, and they behave differently.
The reason can travel. Write it down where your agent reads, and the agent carries it next time without you.
The no comes first, and sometimes it comes alone. You read the draft, something is off, and you can’t say what yet. That feeling is your taste running ahead of your words, and it’s the part no file holds until you find the words for it.
So I don’t believe the flattering version, “AI agents don’t have taste. You do.” It tells you to keep your judgment in your head, which means you have to be in every loop forever, repeating yourself. Here’s the version I teach:
Agents don’t have taste. They have yours, if you wrote it down.
“But I don’t have taste in design”
Or in marketing, or in code.
You’re not a designer, so how would you know good?
You don’t need taste in the field. You need a reader. If you can say “my customer would hate this” or “my team would never read past line two,” that’s the whole skill. It’s also exactly what the agent is missing.
T.A.S.T.E, in five questions
This is the loop for turning your no into something the agent can use. Run it every time you hand work to an agent, and a yes to each question means it’s working.
[T] ell the intent.
Before the how, give the why.
Say who this is for, what done looks like, and which calls the agent can make without checking with you.
A good brief shows in what doesn’t happen next: did it get to work without asking you anything your brief should have answered?
—
[A] llow autonomy.
Point at the destination and leave the road to the agent.
Dictate every step and you get your own idea back, only slower.
So look for its fingerprints: did it make at least one call you didn’t spell out, and can you see which one?
—
[S] ee the work.
Review the work itself, not the agent’s summary of it, and don’t ask the agent whether it’s good.
Hold it up against something the agent didn’t write: the source, a real test, a reader who wasn’t in the room. Before you judge: did you look at anything it didn’t produce?
—
[T] une the system.
When something’s wrong, fix the file, not only this one output.
A correction in the chat fixes today. A correction in the agent’s instructions fixes every time after.
Is every correction you’ve given twice now written where the agent reads?
—
[E] volve the relationship.
Next time, same kind of job: a shorter brief, and more calls the agent makes alone. That’s how you know the loop is turning: is your brief shorter than last time, with work at least as good?
Where you fit
Every time you tune the system, a reason moves out of your head and into a file. That’s why your agent should need you less each time around the loop.
Less, not zero. Three things stay yours, however short the brief gets. Say why, because the agent can find the how and only you know who this is for. Teach once, because the second time you hear yourself say something, it belongs in a file.
And end with your eye, because that’s where the unexplained no shows up. Stay with it until you can name it, then write the name down, and it becomes a rule.
Your job is the no you can’t explain yet.
Everything you can explain belongs in a file.
Try it this week
Pick one job you’ve given an agent more than once. Before you brief it again, open the first brief you wrote for it.
Find one sentence you’ve typed twice and move it somewhere the agent reads every time, like its custom instructions, a project’s instructions or a skill file. That’s your first Tune.
Then watch two numbers for that job: words in your brief, and rounds of fixes before you accept the work. The brief should shrink and the fixes shouldn’t grow. If the brief got shorter and the fixes went up, you didn’t tune anything. You just stopped explaining.
If you want to run this on your own agent with me in the room, the dates are at getagents.today. Or don’t. The exercise works without me.
That’s it for this week.
If this was useful, forward it to someone (a real human) who keeps typing “make it better.”
See you next week ✌️
-- Tom
(the guy teaching a loop, in a loop)
P.S. An earlier issue on the cost of being needed less: Replace yourself first.
P.P.S. I read every reply, the real me.
'Taste is a no with a reason' is a sharper definition than most I've read. I'd add a design leadership angle: the no you can't explain yet is also how a team learns taste from each other, because the person who can't name it is often the one who has seen the pattern before. Writing the reason down does the same job for people as for agents. 'Tune the system' is the step most teams skip, since correcting the output always feels faster than correcting the brief.