Different AI tasks need different levels of freedom. Some need speed and a fixed format. Others need reasoning and context. One model and one prompt for everything doesn't work.
Write Studio addresses this with two layers of separation: model selection and prompt architecture.
Fast models for meta tasks. Strong models for content.
Meta tasks in Write Studio include document naming, option generation, and context memory extraction. Output is short, format is fixed, no deep reasoning required. Haiku handles all of these. Fast and cheap.
Content tasks include Discuss, Suggest, and Inspire. They read documents, draw on user memory, and need real output quality. Sonnet handles these.
When a task needs both speed and quality, I split it into steps.
What's Next has to respond within seconds but also understand the document. One model can't do both well. So I split it into steps. Step one uses Haiku to scan the document and generate direction options with a tuned prompt. Step two fires after the user picks a direction: Sonnet generates the actual content.
The instant feedback from step one buys patience for step two, and the stronger model in step two keeps the output quality high.
Prompt management: templates vs hardcoded
Discuss, Inspire, and Suggest store prompts in JSON template files. Each has professional and creative modes, plus bilingual user templates. These features change often. JSON means editing prompts without touching code.
Everything else is hardcoded in route files. Document naming and memory updates are short and stable. What's Next depends on runtime variables like memory context. Templates would add complexity without adding value.
The split: high iteration and mode branches go to templates. Fixed tasks and runtime-dependent logic stay in code.
One thing that kept getting clearer through this process: a good AI product doesn't need the strongest model. It needs the right model and the right prompt for each task.