Metaprompting is proving to be a very powerful tool that everyone is using now. It’s like coding in 1995, where tools are not fully developed yet. We’re at a new frontier.

— Jared, YC

In this article:

Prompt engineering is becoming a key element in the advancement of artificial intelligence startups. In an episode of the Light Cone podcast, industry experts shared the most innovative applications of techniques like metaprompting — a method that allows large language models (LLMs) to dynamically refine their own prompts. The episode gathered insights from more than a dozen startups at the forefront of AI development, providing practical examples and strategies that are shaping the future of automation.

The Power of Metaprompting

Metaprompting has established itself as an essential technique in prompt engineering. It is an approach that allows LLMs themselves to dynamically refine and improve their prompts through a process called “folding.” This mechanism enables the model to generate optimized versions of instructions based on its previous execution — something crucial for more complex and adaptive tasks.

A notable example is the startup Parahelp, which provides automated customer support for companies like Perplexity and Replet. The company has shared, rarely in the industry, the prompt used in its AI agents — a strategic piece normally protected as an intellectual property asset.

Practical Implementation Examples

During the episode, a detailed example from Parahelp was shared: a six-page prompt that defines the role of the LLM as a manager of a customer service agent. It includes thorough instructions on how to approve or reject tool calls and how to interact with other agents. The level of detail allows the model to operate accurately in complex and integrated scenarios.

Another company mentioned is Tropier, which uses metaprompting to help organizations like Y Combinator understand and debug multi-step prompt flows. The approach allows LLMs to iteratively review their own instructions, promoting continuous improvement cycles.

The Importance of Evals

One of the learnings shared by the interviewed startups is the value of evals — sets of tests, analyses, and parameters used to understand prompt performance. In many cases, they are considered even more valuable than the original prompts themselves.

Parahelp, for example, emphasizes that their evals encapsulate the situational knowledge needed to adapt prompts to specific contexts. This provides a competitive edge as it allows training agents with greater adherence to end users' realities.

The Fundamental Role of Embedded Engineers

The concept of embedded engineers, popularized by Palantir, is an essential practice for the success of modern AI startups. Embedded engineers work alongside clients, understanding their specific needs and customizing software solutions to meet those requirements.

Some of the top Y Combinator companies have adopted this practice, sending engineers rather than salespeople to work directly with clients like the FBI, quickly tailoring solutions for complex problems.

Conclusion: The New Era of AI Startups

We are in a new era of innovation where metaprompting and prompt engineering play crucial roles in the success of AI startups. Companies that manage to integrate these techniques into their workflows are well positioned to stand out in today's competitive market.

In this context, companies like XMACNA stand out by adopting an AI-first approach, combining expertise in process automation with the development of intelligent copilots that empower people, optimize workflows, and amplify results. Operating across industries — from industrial to service sectors, from small businesses to large corporations — XMACNA demonstrates how advanced prompt engineering and data-based customization can be strategic allies in building truly effective solutions. This is the new frontier of applied artificial intelligence: more than technology, intelligence in service of transformation.

For more insights on how AI is transforming the business world, be sure to explore our other articles, such as AI: Yesterday, Today, and Tomorrow - A Comprehensive Perspective and Digital Employees with AI: The Evolution of Chatbots.