The Process Cage

She pulled out a cage. Literally handed the AI a five-step list and walked away. Then she got mad when the work came out cage-shaped.

Here’s the cage:

Step one — research the prospect’s industry. Step two — three problems they’re facing. Step three — connect those problems to our services. Step four — solutions. Step five — the close.

She ran that through Claude. Proposal came back exactly cage-shaped. Five sections, technically correct, the whole thing reading like it’s been touched by a thousand hands and means nothing to anyone. She called me furious. AI can’t do strategic work. I’m done with it.

Yeah. So. She’s wrong about why.

Here’s the thing — the tool can do strategic work. She didn’t ask it to.

The proposal came back generic because she handed the AI a five-step assembly line and asked it to assemble. Right? That’s what assembly lines produce. Generic.

So I asked her. Because I already knew the answer. How did you train the juniors who used to draft these?

She thought about it for a beat. Said she never wrote the steps down for them. Then — actually, she did. Once. Years ago. Senior partners killed the document because the juniors started sounding like the document. So she stopped. Told them what good looked like instead. Sat them in on calls. Gave them three proposals she was proud of and three she wasn’t and told them to figure out the difference.

Right. That’s how you get good work out of a person.

A junior takes the same five-step list and steps over it. Reads the prospect’s last earnings call and notices the CFO said “operational drag” three times. Opens the proposal there. Not on industry research. Reorders. Ad-libs. Because juniors have judgment. And judgment routes around bad procedure.

AI doesn’t have judgment. It has capability. You tell it five steps, it walks five steps. You tell it the order, it walks the order. You tell it what goes in step three, that’s what step three is.

Procedure to a junior is a guideline. Procedure to an AI is a cage.

And the AI sits in the cage and produces cage-shaped work. Does that make sense?

She read me three of the proposals. Interchangeable. Same opener structure, same evidence patterns, same close. Different prospects, same proposal. Because she’d built the same cage for every prospect. So the AI built the same proposal in every cage.

Here’s what she actually needed to write down. Not steps. A destination.

“This proposal should land like a senior strategist wrote it, not a vendor. The prospect should finish reading and feel like we’ve been watching their business specifically. Name three problems they probably haven’t articulated yet. Show how our work solves each one. Close with a next step that’s specific to this company — not ‘let’s talk.’”

Same content. Different shape. Now the AI has room to think. It can open with the earnings call quote, or a market shift, or a customer behavior pattern — whichever best serves the destination. It can deviate from the obvious path when the situation calls for it. It can sound like a person thought about this prospect specifically, because the spec asked it to.

Procedure tells the AI how to work. Spec tells it what to build.

Procedure is a cage. Spec is a workshop.

Procedure produces output exactly as constrained as the procedure. Spec produces output as good as the destination is sharp.

Three ways of saying the same thing. She needed to hear it three times before it landed. And once it landed, she said something I think most operators would say if they sat with it long enough.

“I do this for every junior I’ve ever hired. I just wasn’t doing it for the AI.”

Right. Exactly.

Here’s the thing. You already know how to write destinations. You’ve been doing it your whole career. Make the client feel heard. Sound like a thoughtful partner. Don’t end on a generic close. That’s how you got good work out of people. You didn’t hand them flowcharts. You handed them outcomes and let them figure it out.

You’ve been doing the opposite with AI. Because AI feels like a machine, you wrote it machine instructions. Step one. Step two. Be safe. Be clear. Don’t leave room for misinterpretation.

Right? But machines that need procedures are the dumb kind. Assembly lines. Pour-the-concrete kind. They need procedures because they can’t think.

AI’s the other kind of machine. The kind that thinks. The kind that needs to know what good looks like and gets to figure out how to get there. You wrote it dumb-machine instructions. It dutifully behaved like a dumb machine. Now you’re calling it dumb.

You knew how to do this. You applied it to the wrong category.

So — procedures aren’t the enemy. They’re right for work that shouldn’t need thinking. Routing parts. Filing returns. Closing the books. Anywhere the answer should be the same every time, because the rules are the same every time.

Proposal drafting isn’t that. Neither is client communication. Neither is anything that should sound like a person thought about it. For thinking work, procedures cage the thinking out. That’s not what you want. You want the thinking in.

So you write specs.

Pick one workflow this week where you’ve handed AI a step-by-step. Proposal checklist. Memo template. Estimate process. Anything with numbered steps.

Throw the steps out. Write a spec instead. Two paragraphs, no list. Describe the destination. Describe what good looks like. Add one example of output you’d be proud of, if you have one. Don’t worry whether the spec is “complete.” Worry whether it’s sharp.

Run the same input through both versions. Compare side by side. Procedure output will sound like a template. Spec output will sound like somebody actually thought about the prospect.

If you don’t see the difference, your spec wasn’t sharp enough. Rewrite it. Make the destination more specific. Make “what good looks like” more vivid. Try again.

You’ll know in an afternoon whether it works. It will.

Hard part isn’t the rewrite. Hard part is trusting that letting the AI pick the route produces better work than telling it the route. It does. Because route-choice is where the thinking happens. Procedures kill thinking. Specs require it.

You’ve been writing destinations your whole career. For juniors. For partners. For clients sitting across the table in proposal meetings. You’ve just been giving AI route directions instead of telling it where you’re trying to go.

That stops this week.

Cage isn’t the tool’s limitation. It’s your spec. And your spec’s the one thing in this you actually control.

So. Write the spec. Let it go.

Framework: The Professional Recipe — Ingredient #5 (Output Over Process). Related failure mode: The Process Cage (#5).

Companion pieces: Stop Hiring AI. Start Building It. — the parent principle. It’s Not Excel — the sister piece on misapplied instincts. Both pieces are about old muscle memory misfiring on a new tool.