Writing effective prompts

Your prompt does two jobs. It tells the agent which pages to go looking for, and it tells the agent what to produce from them. A few habits make both jobs easier and lead to more consistent, better-grounded answers. These apply everywhere you write a prompt: a Deep Dive question, a Skill, a Tear Sheet, or a Data Table.

​

How the agent works, briefly

The agent doesn’t get handed a pile of pre-selected passages. It gets the documents you selected and a set of tools, and it works the way you would: list what’s there, search the text for the terms it expects, count hits to decide which documents are worth opening, then read those pages. It can go back and search again after reading something, so a first search that comes up short isn’t the end of it. Two consequences shape everything below:

• Your wording becomes its search terms. The closer your words are to the words printed in the filings, the faster it lands on the right pages.
• It only cites what it read. Every citation points at a page it actually opened, which is why answers are checkable.

​

Use the words you expect to see in the documents

Name the metrics, section titles, instruments and time periods you’re after.

• Instead of “how is their debt”, write “debt service coverage ratio, direct debt outstanding, debt service schedule, fiscal 2024”.
• Where a term varies by issuer or accounting framework, include the variants, for example “statement of net position (balance sheet)” or “DSCR (debt service coverage)”.

​

Name the documents you want to focus on

The agent sees each document’s name, type and publish date before it reads anything, so you can steer it toward the sources you trust by naming them.

• “Use the official statement and the most recent audited financial statements.”
• “Pull the debt service schedule from the 2023 official statement, not the interim figures.”
• “Prefer audited financials over the budget book where both exist.”

In Deep Dive this pairs with the document panel: the panel decides what is in scope at all, and naming documents in the prompt biases the search within whatever is left. If you only want one filing considered, untick the rest rather than relying on the wording.

​

Keep each prompt focused

Shorter, targeted prompts outperform long ones. When a single prompt blends two unrelated analyses, the agent splits its search budget between them and both halves get less attention than either would alone.

• If you want a debt profile and a separate governance review, run them as two prompts.
• If you want five years of one metric, that is one focused analysis and belongs together.
• A good test: if you could hand the two halves of your prompt to two different analysts with no overlap, split it.

There is no character limit, so being specific costs you nothing. Being unfocused does.

​

Follow-ups stay in the session

In a Deep Dive session, the whole conversation travels with each new question. That is what makes follow-ups work: after a debt service coverage answer you can ask “now break that down by fiscal year” and the agent still knows which coverage figure you mean. The trade-off is that an unrelated question asked in the same session drags that earlier context along, which can pull the search toward off-target pages. Treat a session as one line of analysis. When you move to a genuinely new question, start a fresh session. New session with this context in the session menu gives you a clean thread over the same documents.

​

State scope, periods, and definitions

Spell out the boundaries so the agent does not have to guess.

• Scope: what to include or exclude, for example “water and sewer enterprise only, exclude the general fund”.
• Period: which years to cover, for example “fiscal years 2022 to 2024”.
• Definitions: generic metrics differ by sector and accounting framework, so pin them down, for example “DSCR = net revenues available for debt service divided by annual debt service”.

If you don’t define a standard metric, the agent uses a conventional formula for the issuer type and computes it from the disclosed components when the headline figure isn’t stated. That is usually what you want, but your definition always wins, so state it when it matters.

​

Describe the output you want

Tell the agent the shape of the answer. It follows a requested structure closely, and prompts that fix the structure are easier to read and to compare across entities.

• Sections and their order, or the exact table columns.
• Formatting rules, for example “round monetary values to the nearest USD thousand”.
• How to handle gaps, for example “use N/A for missing years rather than dropping the row”.
• What to leave out, for example “one table, no commentary”.

Reusable prompts benefit from especially explicit output shapes because each run should produce comparable answers across entities.

​

Putting it together

A focused, well-scoped prompt might read:

Operating revenues and expenses for fiscal years 2022 to 2024, water and sewer enterprise only, exclude the general fund. Report debt service coverage as net revenues available for debt service divided by annual debt service. Return one table with columns: fiscal year, operating revenue, operating expense, net position, debt service coverage. Round to the nearest USD thousand and use N/A for missing years. No commentary.

​

When an answer comes back thin

Before rewording the prompt, check the cheaper explanations:

• Was the document even in scope? Open the document panel and confirm it’s there and ticked.
• Was the window too narrow? Widen the Years pill. Most gaps are simply outside it.
• Did you use the issuer’s vocabulary? Try the terms that issuer actually prints, including its own section headings.

If a reusable prompt keeps producing the wrong shape, save the corrected framing as a skill so you don’t retype it.