Prompt engineering, once considered a niche technical skill, has rapidly emerged as a foundational capability in working with large language models (LLMs) and generative AI systems. As these systems become more powerful and pervasive, the ability to craft effective prompts—queries or instructions designed to elicit accurate and useful responses—has become a valuable asset in many industries. However, despite its current importance, prompt engineering is likely to become significantly more difficult in the future. Below, we explore the key reasons for this impending complexity.
1. Model Complexity Is Outpacing Human Intuition
As LLMs continue to grow in scale and sophistication, their internal reasoning mechanisms become less transparent and more difficult to predict. Early prompt engineers could develop fairly reliable heuristics because model behaviours followed reasonably consistent patterns. However, as newer models exhibit emergent behaviours—responding differently based on subtle phrasing changes—those rules of thumb become unreliable.
What works for GPT-4 may fail for GPT-5. Prompt engineering, then, becomes an exercise in navigating a black box with shifting boundaries. This increases both the cognitive load on prompt designers and the time needed for trial and error.
2. Fine-Tuning and Personalisation Will Vary by Context
Future models will increasingly be fine-tuned for specific users, industries, or domains. While this will improve performance and relevance, it will also mean that a prompt working well in one environment might perform poorly in another. As organisations move towards deploying proprietary LLMs trained on domain-specific data, prompt engineers will have to understand not only the model’s base architecture, but also its unique training history and customisation layers.
This fragmentation makes it harder to develop universal prompt engineering best practices. Instead, prompts will need to be highly localised and tailored, increasing the complexity of the task.
3. Multimodal Inputs Will Require Multimodal Thinking
Prompt engineering today is still largely text-based. But with the rise of multimodal AI models that process text, images, video, and audio, prompt engineering will require an understanding of how different types of data interact. A good prompt will not only ask the right question in natural language but may also need to include reference images, diagrams, or audio cues.
Designing such prompts requires a new level of creative and technical fluency—essentially becoming an orchestrator of multiple input types, each with their own syntax, constraints, and interpretive nuances.
4. Automation Will Create a Moving Target
Ironically, the very models that require prompt engineering are becoming better at writing prompts themselves. AI-assisted prompt generation tools are already in use, and they will soon dominate the space. But as models learn to optimise their own inputs, human engineers will face the paradox of declining influence: the better the system becomes at adjusting itself, the harder it is for humans to meaningfully intervene.
Furthermore, this leads to a kind of ‘arms race’ where human-generated prompts must compete with machine-generated ones, raising the bar for precision, clarity, and outcome alignment.
5. Evaluation Standards Are Becoming More Complex
Prompt success is no longer defined solely by factual accuracy or grammar. Organisations are evaluating prompts based on alignment with brand voice, regulatory compliance, emotional tone, ethical guidelines, and cultural sensitivity. As the criteria for a “good” output expand, prompt engineering will require multidisciplinary input—drawing from legal, ethical, design, and linguistic domains.
This evolving standard introduces more variables into the prompt development process and reduces the likelihood of success on the first attempt.
6. Security and Adversarial Threats Will Increase
As prompt engineering becomes integral to operational AI systems, it will also become a target. Prompt injection, adversarial prompting, and other forms of manipulation are growing security concerns. Engineers will need to anticipate how prompts could be hijacked or misinterpreted, embedding protective measures without compromising performance.
Balancing usefulness with safety will become an ongoing struggle, further raising the bar for effective prompt design.
Conclusion: From Skill to Discipline
Prompt engineering is currently seen as a tactical skill, but it is likely to evolve into a complex discipline requiring a blend of linguistics, psychology, computer science, and design thinking. As models become more personalised, multimodal, autonomous, and embedded in critical systems, the challenge of engineering meaningful, safe, and consistent prompts will intensify.
In short, prompt engineering will not just get harder—it will become a sophisticated discipline in its own right, requiring new tools, methods, and collaborative approaches to remain effective.
