Asking questions

You can ask the Tablize Agent anything. But some phrasings get you better answers faster. This page is a short guide to the four shapes a question can take and the words that tip the Agent into the right mode.

You don’t have to memorize anything here — the Agent is generous about ambiguity, and it’ll ask you a follow-up if it’s unsure. But if you want one-shot answers, read this.

The four shapes

Every question you ask falls into one of four buckets. The Agent picks the bucket from the verb you use.

Analyze — most common

The default shape. You want to understand something that just happened.

Good phrasings:

• “Why did signups drop on Tuesday?”
• “Show me revenue by product category for the last 30 days.”
• “Break down churn by cohort and pricing plan.”
• “What’s unusual about this week compared to last?”

Watch for:

• Vague pronouns. “Why is it down?” — the Agent doesn’t know what “it” is unless it’s obvious from context. Name the metric.
• Missing timeframe. “How many orders did we get?” — default is “all-time,” which usually isn’t what you want. Say “last week” or “today.”
• Composite questions. “What’s revenue by region and the top SKUs per region?” works, but you’ll get more readable answers by splitting: first revenue, then drill in.

Automate — when you want the answer weekly

Once a good analysis works, turn it into a recurring job.

Good phrasings:

• “Email me this report every Monday.”
• “Rerun this analysis on next month’s data.”
• “Turn this into a reusable script so I can run it per client.”

Watch for:

• Implicit schedules. “Weekly” defaults to Monday 09:00. Say “Friday evening” if you want something else.
• Parameterization. Scripts are most useful when you name the variable: “per client” is better than “on the next dataset.”
• Recipient. If you want the output to land somewhere other than chat, say where: “post this to the #ops Slack channel every Monday.”

Build — when teammates need a tool

Build questions turn an answer into something others can use without asking you.

Good phrasings:

• “Build a CRUD admin for the users table.”
• “Make a dashboard of this for the team.”
• “Let non-technical people explore this data by date and region.”

Watch for:

• Scope. “Build a CRM” is too big. “Build a form to add customers to the CRM table” is actionable.
• Data contract. Name the table or the query. The Agent writes cleaner Apps when it knows which data powers them.
• Who’s using it. “For the team” vs “public shareable link” changes permissions.

Watch — when silence is the goal

Watch questions run silently and only speak up when something moves.

Good phrasings:

• “Ping me if refund rate goes above 3% this week.”
• “Alert me if any sensor reads above 30°C for more than 5 minutes.”
• “Watch this SKU for the next 48 hours — tell me if inventory drops below 20.”

Watch for:

• Threshold precision. “If it spikes” is vague. “If it goes above X” is a rule.
• Time window. “Overnight” is understood — it maps to your workspace timezone, 20:00 to 07:00. Override with explicit hours if you need to.
• Channel. Default is email. Add “on Slack” or “via webhook” if you want something else.

Follow-ups

After the first answer, follow-ups are the fastest path to depth. The Agent remembers the whole session.

Good follow-ups:

• Narrow: “just for paid users” · “only the US region”
• Expand: “include the last 90 days instead” · “add the app revenue too”
• Compare: “same chart for last month” · “how does this compare to Q3?”
• Decompose: “break this down by source” · “what are the top 5 drivers of that drop?”
• Challenge: “is the difference statistically significant?” · “what would I expect if this were random?”

You don’t need to re-state the original question. The Agent carries context across turns until the session ends.

Attaching context

Three ways to add context without typing it:

• Drag a file into the chat — it attaches to the next turn.
• @-mention a table, integration, or previous Report by name. The Agent resolves it.
• Paste a URL — works for public data sources, pasted CSVs from Gist, published GSheets.

When the Agent asks you a question

Occasionally the Agent pushes back with a clarifying question. This usually means:

• Ambiguous metric. “Revenue” might be gross vs net. It’ll ask.
• Ambiguous timeframe. “Last quarter” — fiscal or calendar?
• Destructive hint. If a question might cost a lot of tokens (e.g., “scan every row of a billion-row table to find outliers”), the Agent describes the cost and asks for confirmation.

Answer briefly. “Gross” or “calendar” or “go ahead” is enough.

When the Agent is wrong

It happens. Three moves, in order:

1. Disagree in chat. “That can’t be right — revenue was much higher last week. Recheck.” The Agent re-examines.
2. Show the SQL. Click the tool-call row. You can read and edit the query — fix it there, or tell the Agent what’s wrong with it.
3. Give it a hint. “The orders table has a voided column — exclude rows where it’s true.” The Agent re-runs.

If the underlying data is wrong, that’s a data problem, not an Agent problem. The Agent only knows what the tables contain.

Next steps

• Reading the answer — what the tool-call rows mean, how to trust them.
• The Keep loop — what to do with an answer that’s worth keeping.
• Split view & artifacts — how big outputs get their own panel.