Skip to content
Resources > Product Articles

Prompt Engineering for Financial Research

By Nicole Sheynin, Content Marketing ManagerPublished:
Text input bar displaying "Prompt Engineering Best Practices" over a light blue perspective grid background.

Prompt engineering is the foundation of AI-driven financial research: how well you ask questions to a search platform or LLM model directly translates to how precise and accurate your results will be. Your AI research tool needs to know who the answer is for, what it will be used for, which sources to draw on, over what period, and in what format. When a prompt is loosely or weakly framed, the system has to infer your intent, which can lead to spending too much time asking follow-up questions and trying to filter out the irrelevant outputs. A well-built prompt, on the other hand, can expedite your time from question to deliverable, ensuring the outputs you get are defensible, reliable, and trustworthy.

Below, we walk through seven steps for writing prompts that return precise, cited answers that will actually save you time and enhance your research workflows. We also cover some of the pitfalls that most often result in unsatisfactory outputs, as well as how using AlphaSense for your research helps you avoid those.

Step 1

Be Specific About What Your Need

When starting your prompt, you need to consider the five key elements of successful prompt engineering:

  • Role: Explain your role to the model, so that it understands what you will do with the output
  • Context: Explain the situation, such as a position under review or an upcoming earnings season
  • Objective: Be clear about what you actually need, such as identifying margin pressure signals
  • Examples and specifications: Elucidate what a good answer looks like, such as quoting management’s exact language
  • Instructions: Include any constraints, such as limiting the review to the last two earnings calls

Not every prompt will require all of these five elements, but keeping them in mind when you start will help you reduce ambiguity and cover all your bases.

In AlphaSense’s Generative Search or SuperAnalyst, you can ask the same prompt in two ways:

  • “What’s going on with bank deposits?”
  • “As a credit analyst, summarize how management teams at U.S. regional banks have described deposit costs on earnings calls over the last three quarters. Group the findings by theme and note any company that diverges from its peers.”

The first is considerably weaker than the second and will most likely result in a much less precise and useful output. The stronger version includes the role, context, objective, and specific instructions in a single request.

Step 2

Use Qualifiers to Sharpen Results

When you pull documents by hand, you choose the ticker, the period, and the document type. A prompt should make the same three choices explicit.

  • Timeframe: Generative Search looks at the past 12 months by default. State the range you want (“since Q1 2023,” “in the most recent fiscal year”) rather than relying on a word like “recent.”
  • Sources: Name the document type: earnings call transcripts, 10-K filings, broker research, or expert calls, so that the model doesn’t have to guess.
  • Entities: Ask a full question rather than entering a keyword or company name alone, and use tickers or clearer terms where a name is ambiguous.

In AlphaSense’s Generative Search and SuperAnalyst, you can type @ in the prompt box to add a company, watchlist, or industry directly. To cover several companies, use a watchlist rather than adding each one manually. The + icon opens the Document Selector, and the All Sources dropdown lets you include and exclude content sets, with web search available as an option. Anchoring to an exact entity removes a guess that free-form prompts leave to the system.

Step 3

Match the Framing to the Question

Use factual framing for data, figures, and quotes: “What revenue guidance did [Company] give on its most recent earnings call?” Use analytical framing for interpretations: “How has management’s tone on pricing power shifted over the last three quarters?” Mixing the two in one prompt dilutes both, so if you need both, ask in sequence.

AlphaSense’s Generative Search and SuperAnalyst will automatically pull from our Financial Data and Canalyst if your question is asking for quantitative answers. If your query requires standardized financial fundamentals across broad public company coverage — revenue, EBITDA, net income, cash flow, margins, consensus estimates, valuation multiples, and market data — the tool will use Equity Screener.

Example prompts that will typically pull from Equity Screener:

  • “Compare NVIDIA and AMD revenue growth over the last 3 years.”
  • “What is Tesla’s current EV/EBITDA multiple?”
  • “Show Apple’s gross margin trend since 2020.”
  • “What are the top 10 semiconductor companies by revenue growth?”

When a query requires company-specific segment data or KPIs and specifically references a company, Generative Search and SuperAnalyst will pull from Canalyst.

Example prompts that will typically pull from Canalyst:

  • “Break out Microsoft’s revenue by segment for the past 3 years.”
  • “What is Snowflake’s net revenue retention?”
  • “Compare Lululemon’s US vs international same-store sales growth.”
  • “Using Canalyst data, show Amazon’s AWS revenue growth.”

To specifically prompt GenSearch or SuperAnalyst to pull from Canalyst, reference the segment or KPI by name, or add “according to Canalyst.”

Step 4

Customize the Output

Outputs can come in an array of formats, and you should ask specifically for what you want. Some common options include a table for side-by-side comparisons, a categorized list for themes and risks, a hierarchy for layered findings, and a narrative paragraph for an executive summary. Add structure and length constraints if needed.

In AlphaSense, you can use the Report or Grid button before you ask your question to set the output type. After a response, Build Slides automatically turns key insights into editable PowerPoint slides with each bullet cited, and Download exports a formatted PDF. Copying a response to the clipboard preserves formatting and citations, so your colleagues can check sources and work directly with the output.

Step 5

Iterate and Verify

Once you have your first output, it’s time to iterate and verify. Consider what the output is missing, and ask to change only the element that is off:

  • Too broad: add a qualifier, a time range, or a named source
  • Too narrow: remove a constraint or reframe as an analytical question
  • Wrong format or tone: specify the output explicitly

AlphaSense can improve your prompt for you with our Enhance prompt option. If you want to get granular with the steps the system took, open the Research plan which breaks down each step the tool took to give you the output. Select any inline citation to open the highlighted passage in the source document, and highlight text in the response to ask a follow-up question or verify a claim. When no relevant content exists, Generative Search and SuperAnalyst say so instead of fabricating a claim.

Step 6

Turn Successful Prompts into Agents

Once you ascertain that a prompt is returning the exact output you need, you may wish to create a repeatable agent if this is a recurring workflow.

In AlphaSense, simply select Create Agent at the bottom of a GenSearch response to transform your prompt into a custom agent. Custom agents can run on a daily, weekly, or monthly cadence with results delivered by email. Agents that use Deep Research can only be scheduled weekly.

Ambiguity is more risky in agent prompts than in single queries because errors can propagate across steps. Be explicit about conditions, exceptions, and the format you expect at each stage.

Example of an agent prompt: “Each week, review new broker research on {company}. Flag any estimate or rating change. Return a table with the broker, the change, the date, and the stated reason.”

Step 7

Delegate the Whole Workflow with SuperAnalyst

Where Generative Search answers a question, SuperAnalyst runs research workflows from retrieval through finished deliverables. When writing a prompt for SuperAnalyst, you give it an assignment rather than a question. For example: “Build an earnings preview for [Company]’s next print. Compare what management guided on the last call with current consensus, flag what is likely to have changed, and deliver a short memo and a companion slide deck.”

The assignment needs to carry the same elements as any effective prompt (objective, scope, and format) and name the deliverables. A single task can return more than one artifact, such as a memo and a deck, and each claim always links back to its source. SuperAnalyst also asks clarifying questions instead of guessing, and it will always ask you to approve the research plan at the start.

You can also make assignments repeatable in the following ways:

  • Save the workflow as a Skill. After getting a successful deliverable, ask SuperAnalyst to save it so you can rerun it on other companies. Type / in the chat to see the Skills available to you.
  • Put standing context in memory, not in every prompt. SuperAnalyst can remember things like your coverage universe, the companies you’ve researched, and your formatting and communication preferences. It builds that memory as you work or when you tell it directly, so you don’t have to explain your context in every request. You can also view, edit, or delete what it remembers under Customize > Memory, or simply ask it in plain language. Memory is stored in a private file specific to your environment and is not used to train the underlying model.
  • Schedule it. Describe the task, the scope, the output, and the schedule: “Each Friday, summarize new broker research on the companies in my watchlist, flag rating, price-target, and estimate changes, and update my weekly brief.” Automations can run daily, weekly, monthly, or quarterly, with up to 10 active at a time.

Common Mistakes in Prompt Engineering

Avoid these common pitfalls when writing your prompt:

  • Entering a keyword or company name alone. Ask a complete question that says what you need.
  • Using “recent” without a range. Define the window, or accept the 12-month default knowingly.
  • Stacking questions. When precision matters, ask one complete question at a time.
  • Blending factual and analytical framing. Pull the facts first, then ask for the interpretation.
  • Assuming the system will infer the right document. Use an @ mention or the Document Selector when a specific company, watchlist, industry, or filing matters.
  • Treating an assignment like a question. If the desired output is a deliverable, name the preferred output format explicitly, along with the scope and cadence, or simply hand it over to a SuperAnalyst skill.

Better Prompts, Better Outputs with AlphaSense

AlphaSense makes it easy to engineer the right prompts and get the best possible outputs. Not only is our system designed to take the guesswork out of prompt engineering, but our models pull directly from our extensive universe of premium content sets, from broker research to expert calls to structured financial data, with every insight cited to its source.

And with SuperAnalyst, you can go above and beyond simple question-and-answer prompts and hand off complete assignments: a cited memo, slide deck, or model built from a single request, in a fraction of the time it would have taken you to create.

Start your free trial today.

About the Author
  • Nicole Sheynin

    Nicole Sheynin, Content Marketing Manager

    Fueled by empathy-driven storytelling and good coffee, Nicole is a content marketing specialist at AlphaSense. Previously, she has managed her own website/blog and has written guest posts for various other publications.

Explore more

How to Conduct Precedent Transaction Analysis

Learn how to run precedent transaction analysis in six steps, from building the deal set to the football field, and how SuperAnalyst streamlines each one.
A person in a business suit typing on a laptop with an overlay that reads "SuperAnalyst is working..." next to a loading icon.

How to Conduct Comparable Company Analysis

Learn how to run a comparable company analysis in seven steps, from choosing peers to reading multiples, and how SuperAnalyst automates this workflow.
A person working on a laptop with an overlay text box stating, "SuperAnalyst is working..." next to a loading icon.

Top AI Tools for Financial Analysis

Explore some of the top AI-powered financial research tools on the market today, and discover how to choose the right tool for your business needs.
Isometric illustration of a desktop computer with bar graphs magnified by a blue magnifying glass.

Transform intelligence
into advantage

Develop bold strategies, seize opportunities,
and lead with clarity and confidence.