It’s Good, Try It
Nothing crazy here … just been using Sonnet 5 for a long time to help keep costs down at the $JOB but it quickly became hot garbage.
Sonnet at the high effort level is basically brain dead for my work.
What Do You Do?
Typical day-to-day data engineering work. Airflow Dags so Python, SQL, thinking about ingestion and stuff like that.
After that I do a lot of work in Jira (life of a new engineering manager) and in Confluence writing documentation.
TL;DR
Claude Opus 5.5 at the medium thinking level is really, really good.
That’s basically it.
In Claude set the model to Opus 5.5, set the thinking mode to medium.
Why Medium?
It seems to be the sweet spot around the amount of “turns” that the model will use when it trys to “think” through a problem. At higher levels it seems to over think and burn tokens. This seems to be the sweet spot.
Addendum: CLAUDE.md
You could also consider tweaking your CLAUDE.md to be succinct, to look at your open source projects and libraries and use language servers etc, etc. It’s all online.
But, basically, out-of-the-box it’s fantastic. It’s what Sonnet should be. Let’s see how Sonnet 5.5 (which should come out any day now) performs.