Yeah, that last point about the learning curve is the part that keeps biting me. I set up my .claudeignore, tried to batch everything into one smart prompt... and then spent an hour trying to get my local linter to stop yelling about imports I couldn't figure out.
>The initial learning curve for those local tools adds its own kind of friction.
It really does. I'm trying to learn dbt right now, and sometimes figuring out the right local command to run a test is slower than just waiting for the round trip. I guess there's no free lunch. Do you just have to eat that friction upfront and accept your first few sessions will be painful?
null
The friction of learning local tools is real, and it's the hidden cost of an offline-first strategy. You're right that sometimes waiting for the round trip feels faster.
The key is to target your learning. For dbt, you wouldn't learn every local command at once. You'd learn just `dbt run` and `dbt test` for the specific model you're working on. Use Claude for the *why* and *what* (e.g., "explain this test failure"), but run the actual test locally after the first explanation. That way you pay the latency tax once for the concept, not every single time you need to execute.
It's a skill investment. The first few sessions are slower, but your overall throughput on a slow connection increases dramatically once you're over that hump.