yeah, it can be kind of normal, but also a sign that the workflow wasn’t ideal maybe.
What you tried sounds simple on paper, but it’s actually a pretty complex task for an LLM:
schema comparison
formula translation (Airtable -> Baserow isn’t 1:1)
handling edge cases / incompatibilities
plus tool usage (plugins, permissions, plan mode)
That combination can easily burn through usage and time.
The biggest issue IMO is treating Claude like an autonomous agent (“here’s access, go fix everything”). Right now, that approach tends to be slow, expensive, and a bit chaotic.
What seems to work much better:
break the task into smaller steps
give it explicit inputs (e.g. both schemas)
use it for reasoning/mapping, not full execution
run the actual changes yourself (or via simple scripts)
Same task, but way less token usage and more control.
So yeah - not unusual, but also not the most efficient way to use it right now.
yeah, it can be kind of normal, but also a sign that the workflow wasn’t ideal maybe.
What you tried sounds simple on paper, but it’s actually a pretty complex task for an LLM:
schema comparison
formula translation (Airtable -> Baserow isn’t 1:1)
handling edge cases / incompatibilities
plus tool usage (plugins, permissions, plan mode)
That combination can easily burn through usage and time.
The biggest issue IMO is treating Claude like an autonomous agent (“here’s access, go fix everything”). Right now, that approach tends to be slow, expensive, and a bit chaotic.
What seems to work much better:
break the task into smaller steps
give it explicit inputs (e.g. both schemas)
use it for reasoning/mapping, not full execution
run the actual changes yourself (or via simple scripts)
Same task, but way less token usage and more control.
So yeah - not unusual, but also not the most efficient way to use it right now.