Claude Records A Skill. Research Drops To 30 Minutes. Fragility Remains, Naturally.
David Gewirtz used Claude Cowork's Record-a-Skill to automate a research workflow, cutting a task that previously took hours down to 30 minutes. Claude generated a 19-page document encoding the workflow. However, he identified four drawbacks, including the open question of whether AI automations can survive minor UI changes that traditionally break fragile automation chains.
This illustrates the concept of procedural abstraction versus declarative robustness. Traditional automation encodes rigid if/then chains that shatter on any interface change. AI-mediated automation attempts to encode intent rather than pixel coordinates. The mechanism at play is semantic grounding. The system understands what it is trying to do, not just which buttons to press. Whether that understanding survives real-world interface drift is the question that actually matters.
David Gewirtz at ZDNET used Claude Cowork's Record-a-Skill feature. He documented four specific limitations after successfully compressing a multi-hour research workflow into 30 minutes and generating a 19-page automation document.
- Open Claude.ai and start a new conversation. Describe a repetitive research task you perform regularly, such as gathering summaries on a specific topic. Expected outcome: Claude engages and asks clarifying questions about your workflow.
- Ask Claude to write out a step-by-step procedure for your task as if it were teaching a new employee. Expected outcome: Claude produces a structured document you can review for gaps.
- Test the procedure by feeding it back to Claude in a fresh conversation with a real task. Expected outcome: Claude executes the workflow and you can measure whether the output matches your manual results, revealing where the automation holds and where it breaks.