I was using Sonnet at low effort as a default orchestrator and wanted to reconsider that choice. The orchestrator has a different job from a model carrying out a bounded coding task. It has to keep track of the goal, decide what to delegate, and work out whether the returned results are enough to continue.
I chose to try Opus for that role. I initially picked low effort, then changed Opus and Sol to medium so I could see how the additional reasoning affected my workflow.
That was a preference to test. The benchmark discussion that prompted it couldn’t establish how those effort settings would perform on my own tasks. I didn’t have a matched comparison showing that medium produced better accepted work or reduced the total time spent reviewing it.
Give the reviewer a separate job
For the reviewer, using the model labelled Sol in my setup, I wanted a project review that gave Opus enough detail to act on. I asked for a /scorecard skill that would choose categories appropriate to the project, assess them with evidence, and explain what needed improvement.
A useful report has to distinguish a weak result from missing evidence. If a test hasn’t been run, assigning a confident score doesn’t solve that uncertainty. The recommendations also need enough context for another model to challenge them and turn them into a plan.
The skill was written and checked structurally, then exercised against two fictional scenarios. Its first report correctly kept live-project use marked as untested. That gave me a starting point for independent review without pretending the workflow was already proven.
The routing change had its own small configuration lesson. A follow-up found that one planner entry had no explicit effort value, so the earlier replacement of low with medium had skipped it. Making the value explicit and checking the resolved routes caught that gap. It still didn’t prove that every live session received the requested setting.
I’m interested in model choice because the differences can matter, but I want the comparison to include the whole task. A useful result may need fewer retries, or a more careful review, than the first answer suggests. My setup at this point was an experiment in roles: Opus coordinating the work, Sol providing a separate assessment, and task records gradually becoming good enough to judge whether those choices were helping.