Ask Questions

The Workbench works best when your question is specific enough to anchor the metric, time range, and comparison logic.

The goal is not to “prompt engineer” the model. The goal is to reduce ambiguity in the business request.

Add analysis context from the plus menu

Before asking a question, use the + menu next to the Workbench input when the answer should use a specific file or data source.

The menu gives you three main paths:

• Upload file adds an Excel or CSV file to the current conversation. After upload, Tukun.ai reads the structure, infers tables and fields, and keeps that file available for follow-up questions.
• Connect database starts a new Workbench turn for database setup. You can complete the connection form inside the assistant reply card instead of leaving the Workbench for a separate setup page.
• Select data source opens your common data sources. Pick one to attach it to the current conversation, or use the more entry to go to the full data source list.

The visible source chips in the input show what context will be sent with the next question. Remove a chip if the next turn should not use that source.

The anatomy of a strong question

A strong question usually includes:

• the metric you care about
• the time range
• the grain or cadence
• the breakdown or segment
• the exclusions or filters that matter

Example:

Show monthly revenue by plan for the last 6 months, excluding internal and test accounts.

Start narrow, then expand

Most users get better results from a sequence of small questions than from one oversized prompt.

Recommended pattern:

1. Ask for the base metric and time range.
2. Add the segment breakdown.
3. Add exclusions or special filters.
4. Ask for comparison or interpretation.

Good first-turn questions

• “Show weekly active accounts for the last 12 weeks.”
• “Compare monthly revenue by channel for the last 6 months.”
• “What was conversion rate by landing page last month?”
• “Break down support ticket volume by product area this quarter.”

Questions that need clarification

These are not useless, but they are too broad to trust immediately:

• “How are we doing?”
• “What changed?”
• “Why is revenue down?”
• “What should leadership care about this week?”

Turn them into narrower requests first.

Use follow-ups intentionally

Good follow-ups:

• change the grain
• add or remove a filter
• compare against a previous period
• split a total into segments
• ask which interpretation the system used

Less useful follow-ups:

• repeating the same broad request with stronger wording
• asking for confidence when the definition is still ambiguous
• adding strategic recommendations before the base analysis is right

When prompting is not the real fix

If a question keeps producing unstable answers, the problem may be:

• an unclear metric definition
• multiple conflicting table meanings
• a missing exclusion rule
• a data source that is not yet ready

In those cases, the right next step is not “ask the model harder.” It is usually semantic cleanup or source validation.

A practical test

Before you hit submit, ask yourself:

“Could another teammate read this question and know exactly what result I expect?”

If not, tighten the question first.