A manager wants an employee to identify risks in a new project.
She asks:
Can you confirm that this plan will work?
The employee begins assembling reasons the plan will work.
Now she tries again:
What would make this plan fail?
The project has not changed.
The employee has not changed.
Nobody secretly installed a new risk-management module during lunch.
Only the question changed. Yet a different part of the employee’s knowledge became useful.
The first question prompted confirmation. The second prompted a search for failure modes.
That feels suspiciously like prompt engineering.
AI made the activity easier to notice
We usually describe prompt engineering as the design of an instruction that makes a language model more likely to produce a useful response.
At the base-model level, an LLM produces a response from the interaction between its current context and patterns learned during training. The prompt can specify the task, provide examples, impose constraints and request an output format.
One of the startling results in early large-language-model research was that examples inside the prompt could change task performance without changing the model’s underlying weights. The model could behave as though it had learned the task from context. (Brown et al., 2020)
Even the order of those examples can matter. In a 2022 study of few-shot text-classification prompts, different orderings ranged from near state-of-the-art performance to roughly random guessing. That result concerned particular models and tasks, not every modern LLM, but it makes the general point rather loudly: context is not a neutral container. (Lu et al., 2022)
Humans knew this before computers could autocomplete a sentence.
We called it teaching, interviewing, rhetoric, form design, diplomacy, bedside manner, parenting and—on difficult afternoons—trying to get a toddler into trousers.
The new phrase makes the old activity visible.
The analogy—and its limit
Shared: the literal words are only part of the input. Different: a person can judge the prompter, reject the objective and change the relationship.
A prompt chooses the mental job
Ask someone:
What do you think?
They must first work out what kind of thinking you want.
Ask:
What is the one assumption here most likely to be wrong?
Now the mental operation is clearer.
The second prompt has not added knowledge. It has pointed the searchlight.
This may be the most useful part of the LLM comparison. A good prompt does not merely contain better words. It defines the task.
Are we recalling, comparing, criticizing, predicting, deciding or inventing?
“Is the project on track?” asks for a verdict.
“Which assumption could make the timeline wrong?” asks for a vulnerability.
“Do you agree?” invites a side.
“Which part is supported, and which part is still an assumption?” invites decomposition.
In each case, the answer is partly shaped before the answering begins.
Prompts smuggle in priors
Consider:
Why are teenagers irresponsible?
The question arrives with a small piece of unpaid intellectual labour already completed: teenagers have been declared irresponsible. The respondent’s job is to supply evidence.
Now compare:
Under what conditions do teenagers behave responsibly?
The category has changed. Different memories become relevant.
This is not just wordplay. Tversky and Kahneman famously showed that different presentations of the same decision problem could produce different choices. Their experiments became foundational evidence for framing effects, although the size and reliability of any framing effect depend on the task and context. (Tversky & Kahneman, 1981)
Questions can also affect what people report remembering. In a classic experiment, Loftus and Palmer changed the verb used to ask participants about filmed car collisions. Estimates differed depending on whether the cars had “hit,” “collided” or “smashed,” and later reports of broken glass also varied—even though the films contained no broken glass. (Loftus & Palmer, 1974)
The cautious conclusion is not that wording controls the mind.
It is that a question can participate in constructing the answer it appears merely to collect.
Survey designers know this. Trial lawyers know this. Parents discover it shortly after asking, “Why did you do that?” and receiving a defence brief instead of an explanation.
The human prompt is larger than the sentence
An LLM prompt may include system instructions, conversation history, retrieved documents, examples and formatting rules.
A human prompt also has a context window. It is just harder to see.
It includes:
- who is speaking;
- what happened five minutes earlier;
- whether anyone else is watching;
- how much sleep the receiver had;
- what response is rewarded;
- what disagreement might cost;
- and which button has been made bright blue.
“We should talk” at 10:00 on a calm Saturday is not the same prompt as “We should talk” arriving from your manager at 4:57 on Friday.
The words are identical.
The nervous system has performed some additional formatting.
This is the question underneath my work on Violet: what does a message make noticeable, believable, effortful and important for this receiver, in this moment?
I currently use five handles—Spotlight, Substance, Source, Strain and Stake—not as a formula for controlling people, but as a way to inspect the whole prompt.
A medication instruction may contain perfect substance and still fail because the required action is buried.
A warning may be visible and still fail because its source is not trusted.
A useful request may be ignored because it arrives when the receiver’s attention is already occupied by something more urgent.
There is no perfect message floating free of a receiver.
Prompt quality is relational.
Humans have prompt injection too—sort of
You enter a supermarket intending to buy milk.
The smell of bread, end-of-aisle displays, loyalty points and a sign announcing ONLY TWO LEFT begin issuing rival instructions.
You open social media to reply to one message.
Twenty minutes later, you are learning why a stranger’s dishwasher-loading technique represents the collapse of civilization.
You enter a meeting to evaluate a proposal.
Before discussion begins, the most senior person says, “I’m extremely excited about this.”
The environment has inserted a new objective:
Reach a conclusion that is compatible with continued employment.
Calling this “human prompt injection” is playful, not scientific. A person is not a vulnerable instruction parser. But the analogy reveals something real: our chosen task is constantly competing with instructions embedded in the environment.
How often am I doing what I intended to do?
How often has the room quietly rewritten the task?
Here is where the analogy breaks
If you dislike an LLM’s answer, you can revise the prompt and try again.
Try that repeatedly with a person and the person may revise their opinion of you.
Humans do not respond only to the requested task. We infer why the request is being made.
We ask:
- Are you helping me think or steering me toward your answer?
- Is it safe to tell you the truth?
- Will you use my answer against me?
- Why have you made refusal difficult?
- Do you see me as a collaborator or an obstacle?
An employee may recognize that “What risks do you see?” is ceremonial because the executive has already committed publicly.
A child may hear “Can you tell me what happened?” as an accusation because of the parent’s tone.
A customer may notice that “Confirm your preferences” means “Please surrender more data.”
The receiver models the prompter.
That changes everything.
People have dignity, relationships, bodies, histories, competing commitments and the capacity to resist. We can reject the requested output, challenge the premise or decide that the person asking is no longer trustworthy.
The LLM analogy is useful when it makes communicators more responsible for context.
It becomes dangerous when it makes them see autonomy as a bug.
Better prompting or better manipulation?
Every message frames something. Even a list must have a first item.
So “never influence anyone” is not a workable ethical standard.
A better distinction may be between supporting thought and manufacturing compliance.
A responsible human prompt:
- makes its real purpose reasonably clear;
- includes information needed for judgment;
- does not manufacture urgency or hide alternatives;
- makes the requested mental task understandable;
- allows disagreement without disproportionate punishment;
- and treats the receiver’s interests as relevant.
“What evidence would change your mind?” does not dictate a conclusion. It prompts reflection on what would count as learning.
“Click here immediately to avoid losing your account” uses attention differently—especially when the threat is exaggerated and the alternative is hidden.
Both may be effective.
Effectiveness is not the same thing as legitimacy.
A small pre-send test
Before sending an important message, I want to ask:
-
What mental task am I actually requesting? Recall, comparison, judgment, confession, agreement or action?
-
What assumption have I inserted? Does the question quietly presume guilt, urgency, incompetence or consent?
-
What will the receiver notice first? Is that also the most important thing?
-
What does an honest answer cost them? Time, status, safety, dignity or belonging?
-
Can they meaningfully reject my frame? Or have I designed a multiple-choice question with only one survivable answer?
The best prompt may not produce the answer I wanted.
It may expose that I was asking the wrong question.
What I am still trying to understand
AI made prompt design feel technical because we can alter a sentence and watch the output change within seconds.
With people, the output includes something slower and less visible: trust.
A perfectly engineered request might win the immediate response while teaching the receiver to become guarded next time.
So perhaps we should judge a human prompt twice:
Did it help in this moment?
And what kind of mind—and relationship—did it encourage for the next one?
Until the next strange question,
Osagie