Separate the tasks
A research workflow can include preparing data, proposing a hypothesis, implementing rules, testing and monitoring. A model that drafts code is doing a different job from a model that produces a forecast. Document the inputs and expected output for each task before choosing a model.
Keep rules inspectable
Write down the asset universe, decision time, entry, exit and position limits. Inspect generated rules before a test. Natural-language instructions alone may leave timing or missing-data behavior unspecified.
Evaluate the whole workflow
Evaluate model assistance against a defined task: valid code, reproducible calculations or an explicitly held-out prediction test. A well-written explanation is not evidence of a profitable strategy. This article introduces a workflow; it reports no trading results.