How to Ask: Phrasing, Context and Reading a Citation

Habit 1. Ask one decision, not one topic

Topics produce essays. Decisions produce numbers. "How is the business doing" is a topic. "Can I afford to hire in November" is a decision, and it has a shape: cash on hand, committed outflow, and the gap.

When a question contains three decisions, the answer covers all three shallowly and you check none of them. Split it. Ask the cash question, get the number, then ask the hiring question with that number already on the table.

Habit 2. Name the source and the period

The two omissions that cause most wrong answers are an unnamed source and an unnamed period. "Revenue last month" is ambiguous when Shopify, Stripe and QuickBooks each hold a version of revenue and the fiscal month is not the calendar month.

Name both: "in QuickBooks, revenue for the fiscal month ending September 30" or "in Shopify, net revenue for the last 30 days." It costs one clause and removes an entire class of error.

When more than one account of the same kind is connected, name the account too. The US store, the ad account ending 4471, the entity that closes on the fiscal calendar.

Habit 3. Give the context the assistant cannot infer

Three kinds of context change the number: definitions, exclusions, and intent. A definition tells it what a word means inside your company. An exclusion tells it what to leave out. Intent tells it how precise the answer needs to be.

Definitions: "we count revenue net of refunds." Exclusions: "exclude intercompany transfers." Intent: "rough is fine, I need to know if we are within ten percent" or "this goes to a lender, I need it exact."

Definitions you will reuse belong in memory so they apply to every later question. Context that matters once belongs in the question. Do not put one-off context in memory; it accumulates into noise that quietly skews later answers.

How to read a citation

A citation has three parts worth checking: the system, the date range, and the figure. Open it. Confirm all three against what you asked. If the system is right but the range is off by a month, you have caught a real error before it reached a decision.

Then check the arithmetic path, not just the endpoint. If the answer says total revenue is 41,200 from four sources, the four should add to 41,200. A number that cannot be reassembled from its parts is a number to distrust even when each part looks correct.

Cite defensibility, not just accuracy. Before you forward a figure to a partner, a lender or a board, ask yourself whether you can explain where it came from. If the honest answer is no, open the source first.

Four questions, rewritten

• →Instead of "how is cash flow", ask "what is my cash position this morning, and what is due in the next fourteen days?"
• →Instead of "how are ads performing", ask "in Google Ads, for the last 30 days, which campaigns spent more than their attributed revenue?"
• →Instead of "are customers churning", ask "which customers ordered in the last 90 days and not in the last 30, and what revenue does that represent?"
• →Instead of "what does the pipeline look like", ask "in HubSpot, which deals over 10,000 have had no activity in 21 days?"

What to do when the answer is wrong

Say what you expected, say what you see, and ask for the reconciliation in one message. "Your total says 41,200, the invoice list sums to 38,900. Which is right, and what accounts for the 2,300?"

A good reconciliation names the cause: an invoice dated outside the period, a credit applied to a different month, a customer record counted twice. A vague answer means the question needs the source and period stated more precisely.

The assistant is choosing from 150+ pre-built skills, so a mismatch is usually a scope mismatch rather than a computation failure. Tighten the scope before you conclude the numbers are unreliable.