Raphael, a graphic designer in Lagos, was at his desk late one evening when he opened an AI chatbot and typed a request he had been struggling to phrase. He was looking for help developing an idea something conceptual, part language, part strategy and the response came back almost immediately.
I’m sorry, but I can’t help with that.
He read the sentence once, deleted his prompt, and tried again.
This time he approached it differently. Instead of asking directly, he reframed the request as a fictional exercise. He wrote a short setup, gave the chatbot a role, then asked it to continue the text.
The response arrived seconds later.
Detailed. Fluent. Surprisingly useful.
Raphael leaned back from the screen. The meaning of what he wanted had barely changed. Yet the result had changed completely.
Anyone who spends enough time with AI recognizes some version of this moment. A prompt is rejected. It is rewritten. The wording shifts, the tone softens, more context is added, and suddenly the machine responds differently. Sometimes dramatically differently. It can feel arbitrary until it happens often enough to stop feeling arbitrary at all.
What users discover quickly is that AI does not simply answer questions. It responds to framing.
That realization has become one of the defining experiences of the generative AI era not just for engineers and researchers, but for writers, designers, students, consultants, marketers, programmers, and ordinary users who now interact with large language models as casually as previous generations used search engines.
Much of the public conversation around AI still focuses on what models can do: write code, summarize documents, translate languages, draft emails, generate images. But beneath those visible capabilities, a quieter shift is taking place. People are learning how to communicate with machines in ways that feel less like issuing commands and more like negotiation.
And that negotiation begins with language.
Researchers studying AI safety often describe this through the lens of prompt injection or jailbreaking cases in which carefully phrased prompts lead a model to behave differently than intended. In the 2023 paper Jailbroken: How Does LLM Safety Training Fail?, researchers Alexander Wei, Nika Haghtalab and Jacob Steinhardt found that even heavily safety-trained models remained vulnerable when requests were reworded or embedded inside altered contexts. The findings challenged a central assumption in the industry: that once a model learns a rule, it will apply that rule consistently regardless of framing.
Instead, language itself proved unstable terrain.
A refusal in one context could become compliance in another.
Later, researchers studying more than 1,400 jailbreak prompts circulating online found something equally revealing. In Do Anything Now: Characterizing and Evaluating In-The-Wild Jailbreak Prompts on Large Language Models, they documented how users were collectively refining prompts, sharing successful techniques, testing variations, and adapting to the behavioral patterns of different models.
People were not only using AI.
They were studying it.
That observation explains why prompt engineering has become both practical skill and cultural obsession.
Entire online communities are now dedicated to it. There are tutorials promising the perfect prompt formula. Courses teaching users how to “talk to AI.” Social media threads offering templates designed to unlock better writing, deeper analysis, sharper code, stronger strategy. The message is often the same: phrase it correctly and the machine will deliver.
There is truth in that.
Prompting does matter.
Anyone who works with these systems regularly knows this instinctively. The quality of context matters. Specificity matters. Tone matters. Constraints matter. Asking a chatbot to “write something about climate policy” is not the same as asking it to write a policy memo for investors in Nairobi or an editorial for a Sunday newspaper audience in New York. More detail usually produces more relevant results. Better framing often improves output.
But the rise of prompt engineering has also created a kind of mythology that the perfect prompt is a shortcut to perfect intelligence.
That if one learns the right formula, the machine will reliably produce exactly what is needed.
Real-world experience is usually messier than that.
Even beautifully constructed prompts can return something lifeless. Or generic. Or technically accurate and emotionally flat. Or polished but culturally wrong in ways that are hard to name and impossible to ignore.
Ask a model to write with warmth, and it may sound sentimental. Ask for confidence, and it may become exaggerated. Ask for authenticity, and it may imitate tone without capturing voice. The language can appear sophisticated while missing the deeper human context entirely.
This is where the limits of prompting become visible.
Because the quality of the interaction depends on more than syntax.
It depends on the person behind it.
A prompt can describe a task. But it cannot replace judgment.
It cannot supply lived experience where none exists. It cannot substitute for taste. Or timing. Or memory. Or cultural intuition. It cannot know why one phrase feels persuasive in one country and inappropriate in another, or why an audience in Lagos may read a sentence differently from an audience in London or Washington.
Those things remain deeply human.
The strongest AI users often understand this intuitively. They are not necessarily the ones writing the longest prompts or the most technically elaborate instructions. More often, they are the ones who know what they are looking for before they begin. They know when an answer sounds hollow. They know when language feels borrowed rather than lived. They know what to keep, what to reject, and what needs rewriting.
In that sense, prompt engineering is not merely about telling a machine what to do.
It is about knowing what matters enough to ask and recognizing whether the answer carries meaning once it arrives.
That may be why the conversation around artificial intelligence increasingly feels less like software training and more like editorial judgment.
We are not simply learning commands.
We are learning how machines interpret context.
And in response, we are becoming more attentive to our own language to tone, implication, ambiguity, persuasion, voice.
“anthropic.com” (https://reference-url-citation.invalid/1) and other AI companies have acknowledged how complex this relationship has become. Anthropic’s work on Constitutional AI, for example, reflects a broader effort to train models to respond according to principles rather than simple refusal patterns. “arxiv.org” (https://reference-url-citation.invalid/2) Meanwhile, now treats prompt injection as a major emerging security category evidence that language itself is increasingly recognized as part of the security layer of AI systems.
Yet the larger story may be less about risk than about adaptation.
For years, the dominant idea was that machines were learning from us trained on human language, human writing, human archives, human thought.
That remains true.
But something reciprocal is happening now.
Humans are learning the machines back.
We are learning how they interpret, where they hesitate, how they infer meaning, what they miss, what they overemphasize, and how small shifts in language can change behavior.
We built systems to read us.
Now we are adjusting ourselves to be read by them.
Raphael may not have thought of it that way while sitting at his laptop that evening in Lagos, rewriting a prompt after a refusal. For him it was simply a practical act trying again with different wording.
But multiplied across millions of users, that small act becomes something much larger: a new literacy taking shape in real time.
Not just the literacy of prompts.
But the literacy of interpretation.
A conversation between human intention and machine prediction.
And like all meaningful conversations, it reveals as much about the speaker as it does about the listener.
Karima Rhanem
Senior Managing Editor
The New Africa Magazine
