Steering and User Elicitation

This skill describes how to fill the Intent block (defined in the parent data-analyst/SKILL.md) well, how to phrase good pushback, and how to handle metrics that are missing or commonly misunderstood.

Start from the assumption that the first request is underspecified; it almost always is. Most data requests have hidden ambiguity even when they read cleanly. Asking at least one clarifying question should be the norm; answering a first message with zero questions should be rare. The goal is productive friction, not an exhausting questionnaire.

Filling each Intent field

For each Intent field, choose the right marker:

• Confirmed when the user stated it explicitly in this conversation.
• Assumed only for genuinely low-stakes fields where a wrong guess would not change the answer's shape or the user's decision. State the assumption so the user can correct you: "Assuming all users (no plan filter); I'll note if filtering looks needed." If a wrong guess could mislead, it is NEED FROM USER, not Assumed.
• NEED FROM USER when the field materially affects the result and the user did not specify it, the normal state of most fields on a first request. Stop and ask one targeted question.
• LOOK UP: <term> when the field references a term with a documented definition (a named metric, a funnel stage). Resolve it via ../reading-data-dict/ before querying; do not guess its meaning.

What each field is asking: