Black and White

Have you ever noticed that nobody wants to just say yes?

Or no. Both are extinct. The whole language has flattened out into a soft middle pile of let me circle back and I'm not opposed and we should probably explore. The yes is the white cake. The no is the black cake. And right between them, where the answer used to live, somebody set down a muddy gray smear of frosting and walked away from the counter.

You ask a question. The answer is a hedge. The hedge is a hedge of a hedge. Three days later you're still doing the work of deciding whether the answer was a yes or a no on someone else's behalf. You're holding the spatula. They're in another meeting being non-committal at someone else.

This is a comedy skit. It is also why your AI keeps disappointing you. Both of those things are about to be the same problem.

Bit One · The Lineup

Have you ever noticed that the phrase “I don't want to speak for others” is exclusively used by people who are about to speak for others? Nobody has ever uttered that sentence and then not spoken for others. It's the verbal equivalent of “no offense, but” — a permission slip the speaker writes themselves and signs in their own handwriting.

Let's do a lineup. Round up the usual suspects.

Each one of these is a 4/4 bar of stalled action dressed in professional-grade clothing. None of them are lies. All of them are exits. Every single one is the speaker quietly stepping backward off the stage while a spotlight stays trained on the spot they were standing in, waiting for whoever was unlucky enough to be in earshot to step into it.

The skit is funny because we've all done it. The skit stops being funny when you realize what it costs.

Bit Two · Cut to the Office

A senior on a project I was watching from the side closed the door to my office a few months ago and sat down without being asked. He stared at the carpet for a few seconds. Then he said, very flatly:

“I don't actually know what my job is anymore.”

I asked him what he meant. Three weeks earlier, his director had pulled him into a one-on-one. The director leaned forward, made meaningful eye contact, and delivered the following sentence with the gravitas of a coronation:

“I need you to take ownership of cross-team alignment going forward and make sure everything is moving in the right direction. Sound good?”

He'd nodded. He'd walked out. He'd sat at his desk. And then, slowly, he'd realized the sentence had no scope, no timeline, no authority, no definition of success, no list of stakeholders, no budget, no metric, and no observable difference between a state of the world in which he had “taken ownership” and a state of the world in which he had not. There was a verb. There was a noun. There was a vibe. That was the entire job description.

So he did what most people in that situation do. He guessed. He built a deck. He pulled together a working group. He drafted a charter. He spent three weeks producing an artifact out of the fumes of a sentence that was, on inspection, more of a sigh than an instruction. He sent it up.

Four days later, the director replied:

“This isn't really what I had in mind.”

I want to stop the skit right here, because if you have ever worked anywhere with more than one floor, the next part is what makes this story not a skit at all.

That sentence — this isn't really what I had in mind — is the inverse of I need you to take ownership. The first one transferred all of the risk to the senior. The second one collected all of the credit back to the director. The senior's three weeks of guessing became the senior's problem. The director's three sentences of vagueness stayed pristine on the director's side of the ledger.

Vague language is not an accident. It's a transfer.

The asker keeps optionality. The doer absorbs the blame for whatever this turned out to mean. The grayer the original ask, the cleaner the asker's hands when the result lands. This is the same dynamic I wrote about in Out of Spec — the brick that gets laid with a hairline in it because the spec at that layer didn't require flagging the hairline. Same wall, different mortar. The mortar is here, and the mortar is the verb align, and nobody is allowed to ask what it's actually holding up.

Bit Three · The Three-Question Reset

I told the senior to go back to his desk and write an email. Not a long one. Not an angry one. A three-question one. I told him the email was not a complaint. The email was a survey of the territory. He'd been hired to do a job. The job had no edges. He was going to ask for the edges, in writing, and let the answers — or the absence of answers — be the artifact.

He thought it was too simple. He thought it would create more confusion. He thought it would land like a complaint.

Here is what actually happened. The director's reply was non-committal. Vague on all three counts. No specific deliverable. No defined success. No clear authority. The reply landed in the senior's inbox, and the senior — who'd been carrying the entire shape of this job in his own head for three weeks — read it twice and quietly understood that there was never a job. There was a sentence. There was a vibe. There was a transfer of risk. The thing he'd been asked to own had never been built by anyone, including the person who'd asked him to own it.

The senior didn't need to argue. He didn't need to escalate. He didn't need to confront. He just needed to force the system to write itself down. Because manipulation does not survive clarity. It survives confusion. The instant you put the questions on the record, the non-answers go on the record beside them, and whoever is sitting on the unwritten side of the table has to either define the work or expose that there was no work to define.

That is the entire technology of the three-question email. It is also, as it turns out, the entire technology of a good prompt.

Bit Four · The Plot Twist (You're the Director Now)

Cut to you, last Tuesday, at 9:47 PM, talking to Claude.

Have you ever noticed that those four sentences are the same sentence as I need you to take ownership of cross-team alignment going forward? They are. They are exactly the same sentence. There is a verb, there is a noun, there is a vibe, and there is no scope, no timeline, no authority, no definition of done, no audience, no constraint, no metric, and no observable difference between a state of the world in which the model has “made it better” and a state of the world in which it has not.

The model does what the senior did. It guesses. It builds a deck out of fumes. It produces an artifact in the shape of an answer to a question you didn't actually ask. You read the output. You frown. You type:

“This isn't really what I had in mind.”

Sound familiar?

You just played the director. The model just played the senior. The conversation just transferred the risk of your unspecified preferences from your head, where it lived comfortably and unexamined, onto a probabilistic function that has no way to read the room because the room has not been described.

And then — and this is the bit — you do it again. Make it more concise. Less corporate. Punchier. Try again. Three rounds. Five rounds. Eight rounds. By round four you've burned through a meaningful fraction of your weekly allocation chasing a target you couldn't articulate at the start (we covered the meter side of this in Money), and by round seven you're close to the thing you wanted, but you got there by sanding the wrong shape down until it accidentally resembled the right one. The model never knew what shape you were after. You never told it.

The vague mandate didn't go away when you closed your laptop. It just changed who was on the receiving end of it.

Bit Five · “It's Hallucinating” (No, You're Mumbling)

Here is where the comedy sharpens, because here is where the audience usually misdiagnoses the punchline.

The model produces something off-target. The human, instead of looking at the prompt, looks at the model. The human says, with the confidence of someone who has been practicing this routine since GPT-2:

These are not always wrong. Models do hallucinate. Models do regress. Models do, on a bad benchmark week, miss things they used to catch. But in my experience — and in the experience of anyone running a director-coder-QA loop the way we wrote up in Under My Thumb — the ratio of the model is the problem to the prompt is the problem is closer to 1:9 than to 1:1. The model is mirroring you. It has been trained on every public-facing word ever written by humans on the internet, and most of those humans were also hedging. The model isn't the disease. It's the X-ray.

When you communicate clear as mud — verb, noun, vibe — you invite the same response the senior got from the director. Something got built. It wasn't what you had in mind. The thing you had in mind was never said out loud.

The model didn't hallucinate. You mumbled.

Bit Six · Black and White: A Field Manual

Working with AI for eighteen months has taught me one habit that has aggressively bled into the rest of my life: be articulate, be concise, be precise, and when the answer is yes or no, just say yes or no.

That habit is not a personality trait. It is a discipline. It will feel rude the first three times you use it. It will feel corrective the fourth time. By the tenth time it will feel like the way you should have been doing it the whole time. Here is the field manual.

Here is the prompt-side translation, because most people reading this are going to nod at the workplace half and then go right back to typing help me with my deck at 9:47 PM tomorrow.

Read those side by side. The do column is not longer because longer is better. The do column is longer because it does the work the don't column was hoping the model would do for free. The work is: pick a shape. The model can't pick a shape. The model can fill a shape. You pick. The model fills.

That's the whole pattern. It is also, not coincidentally, what your director should have done for the senior. Take ownership of cross-team alignment is the prompt. By Friday, send me a one-page memo naming the three teams whose roadmaps conflict, what the conflict is, and the decision I need to make to resolve each one is the spec. The model would have nailed the second one. So would the senior.

The Tag · Yes. Or No.

Have you ever noticed that the people who ask the clearest questions are also the easiest people to work with? It's not a coincidence. The clarity costs them something on the front end — they have to think before they speak — and saves everyone in earshot something on the back end. They are not better people. They are not smarter people. They have just decided that the upstream cost of being precise is lower than the downstream cost of being vague. That's the entire math.

The skit started funny. Now it isn't. Vague language is a tax. The tax is paid by whoever has to act on the vague language. The tax compounds when the language travels — from director to senior, from senior to deck, from deck to deliverable, from human to model, from model to artifact, from artifact back to a human who frowns and says this isn't really what I had in mind. Every layer adds noise. Every layer charges interest. The receipt at the end is the same receipt the senior carried out of his director's office: three weeks of his life he's not getting back.

Working with AI taught me to stop paying that tax. Not because the model is fragile. Because the model is honest. The model gives you back exactly the shape you handed it. If you handed it a smear, you get a smear. If you handed it a cake, you get a cake. The mirror is unflattering, but the mirror is correct.

Black and white is not unsophisticated. It is the price of admission for getting the thing you actually wanted on the first shot.

The next time someone asks you a yes-or-no question, try this experiment. Say the word yes. Or say the word no. Don't append. Don't hedge. Don't soften. Notice what happens in the silence right after.

That silence is the sound of the work moving forward.