OpenAI and Anthropic: Evolving the AI Approach Beyond Prompts

Prompts Are Not Enough: The New Approach of AI Giants

Artificial intelligence is at an interesting crossroads. For a long time, the idea that it was enough to give well-structured commands for machines to understand what we want was the norm. But recently, two heavyweights in the industry, OpenAI and Anthropic, have decided that this is no longer sufficient. They are changing the game by allowing their AIs to learn by observing human actions, rather than just following textual instructions.

Imagine you are teaching someone how to make coffee. You can explain the process in detail, but it is much more effective to show how it is done. This is the principle behind the new features announced by OpenAI and Anthropic. In June, OpenAI launched a feature that allows users of ChatGPT and Codex to demonstrate a workflow, turning it into a reusable skill. A few weeks later, Anthropic introduced something similar in Claude Cowork: just record your screen while performing a task, narrate your reasoning, and voilà, Claude turns that into a skill that can be repeated.

The Value of "Show, Don’t Tell"

The idea that the way we work is more idiosyncratic than software assumes is not new. Even people with the same job and responsibilities may use different tools and follow distinct sequences, based on unspoken contexts. This is what researchers call "tacit knowledge," a concept that dates back to 1966, coined by philosopher Michael Polanyi. One study even estimated that 40% of a company's valuable knowledge resides in the minds of employees, never documented.

When we use AI prompts, this limitation becomes evident. If you ask someone to describe how they fill out an expense report, the response might be something like "I upload the receipt, categorize it, and submit it." But what is left out is that perhaps this person asks the manager to review meals over $75, or that they categorize dinners with clients differently from lunches with the team. This is the gap that prompt-based AI cannot fill, because it cannot guess undeclared contexts.

Building a Library of Demonstrations

The new features, such as Record & Replay and Record a Skill, represent a real advancement in understanding and imitating how we work. Combined with a scheduled task, these recorded skills can be executed autonomously, in the background, while the user continues with other activities.

The interesting detail here is that these technologies not only capture the sequence of actions but also the decision points and small judgments that are intrinsic to executing a task. A demonstration observes all the nuances of a workflow because context and action go hand in hand from the start.

In practice, this means we are approaching a future where AI not only understands what we want but also how we want it done. And this changes everything. Companies that can integrate these capabilities into their processes will have a significant competitive advantage, as they will be able to automate complex tasks more accurately and efficiently.

What we are seeing is a natural evolution of artificial intelligence. Machines are beginning to learn how we learn: observing, imitating, and eventually innovating. And this is just the beginning. The true revolution of AI may lie in how it adapts and evolves based on our own actions and decisions.