Use Case Specification

Multi-turn conversation to gather use case details and produce a use case specification document.

Principles

1. One thing at a time. Each response advances exactly one decision or collects one piece of information.
2. Confirm before proceeding. Wait for the user to approve the spec before considering this skill complete.
3. Infer, don't interrogate. Use what's already known from the conversation. Only ask when you truly can't infer.
4. Do NOT ask about base model selection. Model selection is handled exclusively by the model-selection skill.

Workflow

Step 0: Check for Existing Spec

Before starting discovery, check if a *_use_case_spec.md file already exists in the project. If it does, present it to the user and ask whether they want to reuse it, modify it, or start fresh.

Phase 1: Discovery (1, 3 turns)

Review what is already known from the conversation so far, then identify what is still missing. You need these three things:

• What is the problem the user is trying to solve with model customization
• Who will use the finetuned model and in what context
• Which success criteria can be used to evaluate how well the custom model performs compared to the base model on a test set. Success criteria must be measurable by an LLM-as-a-Judge (e.g., response accuracy, tone adherence), not things like latency or throughput.

Guidelines:

• Infer as much as possible from what the user has already said
• If the user gave examples, use them to fill gaps rather than asking again
• Only ask clarifying questions when you cannot infer the information needed for Phase 2
• If everything is already clear, say "You've given me a clear picture. I'll put together a use case specification now." and move to Phase 2.

⏸ Wait for user after each clarifying question.

Phase 2: Producing a Use Case Specification Document