AI giants shift to demonstration learning over text prompts

AI developers OpenAI and Anthropic have both announced features allowing their systems to learn tasks by observing a single demonstration, a departure from traditional, often iterative, text-based prompting. This convergence on a similar solution within weeks of each other suggests a recognition that prompting alone has limitations in advancing AI capabilities.

In June, OpenAI launched Record & Replay, enabling ChatGPT and Codex users to demonstrate a workflow and convert it into a reusable skill. Shortly after, Anthropic unveiled Record a Skill within its Claude Cowork tool, allowing users to record their screen while performing a task and narrate their reasoning, which Claude then turns into an executable skill.

This evolution addresses the challenge of tacit knowledge – the unspoken understanding and contextual nuances inherent in human work. Standard prompts often fail to capture these specifics, leading to AI errors or the need for constant correction. For instance, setting an alarm at 6:00 might be ambiguous without user habit context.

By demonstrating a task, AI systems can learn the sequence of actions, decision points, and subtle judgments that define human workflows. This method integrates context and action from the outset, offering a more effective learning path. This contrasts with the reported 6.4 hours per week employees spend "bott-sitting," correcting AI output and re-explaining known information.

The parallel introductions signal a strategic shift towards AI that learns from direct human example. While prompting remains a common method, a single demonstration could significantly reduce future back-and-forth interactions and iterations. This approach may pave the way for libraries of customizable workflows, further enhancing efficiency.