Lesson 3 of 14 · Part one: Frame it

Say the Want, Not Just the Rule.

‌‌‌‌​​‌‌​‍‌​‌​‌‌‍‍‌​‍‌‌​​‌​​‌‌​‍Rules collide. “Watch the meter” one day and “burn everything” the next leave your agent guessing. Tell it what you want to happen, and most clashes settle themselves.

9 min to read15 min to tryWorks with any agent
  1. One dayWatch the meter
  2. Another dayBurn everything
  3. What I meantNever run out before the handoff

01 · The conversation

Explicit desires.

One Thursday morning in November my advisor and I were working out why my agents kept losing track of things. Something it had said earlier stuck with me, and I worked it out out loud. I dictate, so this is one long breath:

Mike10:18 AM

OK, regarding the memory management thing, something you said earlier I think there's something I could be doing better. Instead of giving explicit instructions, I think I should be giving explicit desires, like what I want to occur, what my goal is, what I want the outcome to be. When I first started using AI, that's how I did it, and I got nothing I wanted ever, so you trained me to be hyper specific, but now I believe that that's in the way, because on some days I'll say watch the meter, and on other days I'll say burn everything, but what I mean is burn everything all the time, only watch the meters so that you don't end up running out of time to create a handoff, that's the only thing I care about is that I get a handoff from every session, and sometimes there's 50

Fifty is how many agent sessions I sometimes have running at once. One sentence of why turned two rules that looked like opposites into one.

02 · Why it matters

Hyper-specific got in the way.

When I started with AI, I described what I wanted and got nothing I wanted. So I learned to spell out every step. That worked for a while. But exact steps are brittle. Each one is right in the situation I pictured when I said it, and wrong in the one I didn't.

Here's how it played out. Some days I'd tell my agents to watch the meter, meaning how much of the week's allowance they'd used. Other days I'd tell them to burn everything. Both were rules. Neither said what I was after. So an agent holding both could only guess which one won today.

What I was after was simple: every agent leaves a handoff for the next one, so nothing is lost when it stops. Once that's said, the two rules stop fighting. On my plan, the work can run freely, as long as the meter always leaves room to write the handoff. And any one big job still gets one glass, not the keg.

03 · What it means

One reason per rule.

Course 1 has the seed of this. In Lesson 16, Clever, Not Sneaky: “A rule makes more sense to an agent (and to you) once you say what it’s guarding.” And Lesson 2, Say the Goal Out Loud gives the whole project a goal. This lesson puts a reason beside every standing rule, so that when two of them collide, your agent can settle it by what you want instead of by whichever rule it read last.

  1. The rule.

    What to do. “Back up the file before you change it.”

  2. The want.

    What it protects. “Because I can't afford to lose a client's work.”

  3. The clash.

    When two rules pull apart, the agent follows the want, and tells you it did.

One limit, straight from Lesson 8, Never Guess. Settling a clash on its own is fine for cheap things that can be undone. If the clash touches money, something being sent, or anything that can't be undone, the agent still stops and asks you.

And a note on my example. My studio is on a plan with a fixed weekly allowance, so burning it costs me nothing extra. If you pay by the job, your want might be the opposite one. That's the point: the want is yours to say.

The one rule

A rule says what to do. Your reason says what to do when the rule doesn't fit.

Write down what you're after beside every standing rule. When two rules clash, the agent follows what you're after and tells you it did.

When the clash touches money, a send, or anything that can't be undone, it asks you first.

04 · Try it today

Give every rule its reason.

Ten minutes to write the reasons, five to read them back. The reading back is where it pays.

  1. Ask your agent for the list.

    Use the chat where your standing rules live, or paste your custom instructions into a new chat.

    For each standing rule I've given you, write one line beside it saying what I'm really after with it. Show me the list so I can correct the reasons. From now on, when two of my rules seem to clash, follow what I'm after and tell me you did. If the clash touches money, sending something, or anything that can't be undone, ask me first.
  2. Read the reasons back.

    You'll find at least one reason your agent got wrong. That's a rule it has been following for the wrong reason. Fix the reason, not the rule.

  3. Add “because” to the next rule you give.

    Every new rule gets its reason in the same sentence. It costs you five words.

What to take with you

A rule tells the agent what to do. Your reason tells it what to do when the rule doesn't fit.

Where this comes from in course 1. Course 1's Lesson 16 (say what a rule guards), Lesson 2 (the goal for the project) and Lesson 8 (ask when a guess could become a decision). New here: a reason beside every standing rule, and settling clashes by it.

Next · Lesson 4

Don't Let It Shrink the Ask

You asked for the whole thing and got a sample. The smaller version always comes with a good reason. Put your words beside it and the gap shows.

Read Lesson 4

← Previous: Lesson 2, Reasoning Isn't Judgment

Every exchange on this page really happened and is quoted exactly, slips included; […] marks where I trimmed. Times are Eastern. Names of clients, people, companies and products are left out. My agents are AI models; the names are left out on purpose.