Claude Records A Skill. It Writes 19 Pages Of Instructions. Fragility Persists, Naturally.
David Gewirtz used Claude Cowork's Record-a-Skill feature to automate a research workflow, reducing a task that previously took hours down to 30 minutes. Claude generated a 19-page document describing the automation steps. Gewirtz identified four drawbacks and raised the open question of whether AI automations can dynamically adapt to minor UI changes, a problem that has historically destroyed traditional pixel-reading and if/then automation setups.
This illustrates the principle of declarative versus procedural automation. Traditional automation encodes fragile steps: do this, then this, then this. AI automation encodes intent: achieve this outcome. The mechanism matters because intent-based systems can theoretically reroute around a changed button or renamed menu. Whether Claude actually achieves that resilience remains unproven, but the architectural shift from brittle scripts to adaptive reasoning is the real innovation here.
David Gewirtz at ZDNET tested Claude Cowork's Record-a-Skill. He cut research time from hours to 30 minutes but documented four limitations. The 19-page skill document Claude produced suggests the system is doing something more sophisticated than recording clicks.
- Open Claude.ai and describe a repetitive research task you perform regularly, such as gathering summaries on a specific topic. Ask Claude to break it into steps. Expected outcome: Claude produces a structured workflow outline.
- Refine the workflow by telling Claude which steps you want automated and asking it to write detailed instructions for each. Expected outcome: You receive a document resembling the 19-page skill spec Gewirtz described.
- Test the workflow by pasting your task into Claude and asking it to execute each step in sequence. Expected outcome: Claude completes the research in minutes rather than hours, though you should note where it stumbles for future refinement.