Writing Better Prompts

The quality of what you get from Ana depends heavily on how you ask. This is not about learning special syntax — it is about communicating the way you would with a sharp analyst who knows your data but does not yet know what you need right now.

---

​

Always include a timeframe

If you don’t specify a date range, Ana will pull all available history — which could be years of data and is rarely what you want.

---

​

Start specific, then broaden

Vague questions produce vague answers. Before you type, ask yourself: what table does this live in, what time range am I interested in, and what would a good answer actually look like? Answering those three questions in your prompt eliminates most back-and-forth. The specific version tells Ana the table, time period, granularity, and what you are actually looking for. You get a result on the first try instead of a clarifying question.

---

​

Name your metrics

If your company has a specific definition for a term — “active user,” “churn,” “GMV” — include it in the prompt. Without one, Ana will make a reasonable guess that may not match your internal standard.

“Calculate churn rate using the subscriptions table — define churn as any subscription with status = 'cancelled' in the last 90 days”

For definitions that apply to every chat, ask your admin to add them to the context library. Ana will use them automatically.

---

​

Break complex asks into steps

A single prompt asking for five things at once often produces a result that is partially right on all five rather than fully right on any of them. Build up the analysis in stages — each step is a checkpoint where you can catch problems before they propagate. This is especially useful when working with a new dataset or a table you haven’t queried before.

---

​

Use follow-ups, not mega-prompts

Ana remembers everything in the current thread — you don’t need to re-explain context in every message. Once Ana has produced a result, follow-up prompts that reference it are faster and more precise than rewriting everything from scratch.

• “Now filter that to only include accounts created after January 1.”
• “Can you add a column showing the percentage of total?”
• “Re-run this but group by country instead of region.”
• “The Northeast numbers look off — can you show me the underlying rows?”

Think of it as a conversation, not a form submission.

---

​

Ask for charts to verify results

Numbers in a table can hide patterns that a chart makes immediately obvious — especially with time series data, distributions, or anything where outliers matter.

“Plot that as a line chart by week. I want to see if there are any spikes or drops that stand out.”

Charts also help catch data quality issues early. A sudden cliff in March might be a real business event, or a gap in the data pipeline — either way, you want to know before putting the number in a slide.

Instead of: “Is our data complete?” Try: “Plot daily row counts from the events table over the past 90 days. Flag any days where the count drops below 80% of the 7-day rolling average.”

Python handles all chart rendering — just describe what you want.

---

​

Specify the table when you know it

Ana can often infer which table to use from context, but if you know where the data lives, say so. This removes ambiguity and ensures you are looking at the right source of truth — especially in environments where similar data exists in multiple places.

“How many new users signed up last month? Use the accounts table, filter to created_at between March 1 and March 31, 2026, and exclude test accounts where is_internal = true.”

Not sure which table to use? Ask: “What tables contain user signup data?” Ana will tell you what’s available, then you can pick the right one before running the analysis.

---

​

Know when to start a new thread

Follow-ups work well within a single analysis. But long threads accumulate context that can subtly influence later answers — if you asked about revenue in the first half of a thread and then pivot to support tickets, Ana may carry forward assumptions (like the same date filter or customer segment) that no longer apply. Start a new thread when:

• You’re switching to a different dataset or business domain
• You want a clean slate without prior filters or assumptions carrying over
• You hit a sandbox error — a fresh thread resolves this immediately (see Chat Errors)

Stay in the current thread when:

• You’re drilling into the same dataset
• You want Ana to remember a definition or filter you established earlier
• You’re iterating toward a final version of the same analysis

---

​

Ask Ana to explain her methodology

Before sharing a result with a stakeholder, ask:

“Explain in plain language how you calculated this.”

One extra message. Makes the output much easier to defend.

---

​

Turn working prompts into Playbooks

Once you’ve got a prompt pattern that works, you shouldn’t have to run it manually every week. Playbooks let you schedule any analysis to run automatically and deliver results to Slack or email.

---

​

Quick reference