Blog · · Ashutosh Agarwal
The six-part formula for a better prompt
Your AI is only as good as the context you give it. The analyst's guide to prompts that produce genuinely useful investment research: a six-part formula, a full worked example on Apple, and a checklist you can reuse on Monday morning.
TL;DR
- A vague prompt does not produce a vague answer. It produces a confident, generic answer that you then have to redo yourself
- Length is not the variable that matters. Whether the prompt transfers your analytical intent is
- Six ingredients make a research prompt work: audience, objective, scope, analytical lens, output and evidence standard
- We take one 700-word institutional prompt on Apple apart line by line so you can steal the structure
Matterfact Weekend Education Series · Edition 01
Welcome to the first edition of the Matterfact Weekend Education Series. Every weekend we take one practical skill and break it down for analysts and portfolio managers who want to get more out of AI. No lengthy theory, and no need to become an AI engineer. Just techniques you can use on Monday morning.
We are starting with the one that everything else rests on: how to write a good prompt.
Sam Altman put it this way:
id: https://x.com/sama/status/1627796054040285184
caption: Writing a great prompt is an early form of programming in natural language. For analysts, it is also a way to make investment judgment explicit.
Prefer slides? The whole argument is below as a deck you can present or download.
id: prompting
AI cannot read your mind
You have probably heard some version of this line:
You are only as good as your prompt.
src: /assets/images/blog/average_familiarity_by_experts.png
alt: xkcd comic. Two geochemists agree the average person probably knows the formulas for olivine, one or two feldspars, and quartz, of course. Caption: even when they are trying to compensate for it, experts in anything wildly overestimate the average person's familiarity with their field.
side: right
width: 300
It has become a cliché for a reason. A model can read more than you ever will, but it cannot infer the judgment you did not write down. Every gap you leave gets filled with a default, and the defaults are rarely the ones you would have picked.
The same blind spot applies here. You have been living inside the thesis for a week; the model is meeting it for the first time, in one sentence. Everything obvious to you has to be said out loud, because the part you would never think to explain is exactly the part it cannot recover.
label: What the model has to guess
tone: verdict
left: What only you know
right: What it defaults to instead
How deep the analysis should go :: A safe middle depth. Thorough nowhere in particular.
Which aspects of the company matter to you :: Equal weight to everything it can find
What you already understand :: Definitions you did not need
What decision you are trying to make :: No decision, so a summary instead
Which time periods should be compared :: The most recent figures it can source
What your PM will ask next :: Nothing. It answers the question asked, and stops.
What your firm treats as credible evidence :: Whatever it found, with facts and inference blended
Consider these prompts:
Is Apple a good stock to buy?
Bring me up to speed on Apple.
Build me a valuation model for Apple.
None of these is a bad question. Each one would come back with something that looks perfectly respectable. The trouble is that they hand almost every real choice to the model.
Take the first one. What does "good" mean here, over what horizon, and against whose expectations? Should the answer lead with earnings revisions, valuation, the product cycle, regulation, or how the stock is owned? The third prompt has the same problem: a DCF, a multiple, and a sum-of-the-parts are three different pieces of work, and the answer also depends on the forecast period, the drivers, the peer set, and whether you want consensus estimates or your own.
The model has no way to know any of that, so it guesses. When it guesses, it plays it safe, and safe means a broad, competent and forgettable summary.
variant: loop
label: Exhibit 1 · The shallow-prompt loop
loopLabel: and round it goes again
Vague request :: "Bring me up to speed on Apple."
AI fills the gaps :: Horizon, peer set and evidence bar, all quietly guessed
Generic analysis :: Accurate, broad, and answering nobody's actual question
You redo the work :: The hour you saved plus the hour you spent reading it
Prompting is a high-leverage skill
That is the whole test. Not length, not vocabulary, not how technical it sounds. A prompt works when it closes the gaps above on purpose rather than leaving them to the model.
label: What a useful prompt states
tone: verdict
left: What you write down
right: What it stops the model doing
Who the work is for :: Explaining things your reader already knows
What the reader needs to understand :: Answering a different question well
Which questions actually matter :: Spreading equal attention over everything
How the problem should be decomposed :: Describing an outcome without its drivers
What evidence should support the conclusions :: Blending reported facts with inference
What the finished work should look like :: Handing back a shape you cannot use
Here is what that looks like in practice.
From a request to a research assignment
Instead of saying:
Bring me up to speed on Apple.
Try this:
id: the-prompt
emphasis: true
label: Institutional-grade Apple primer
Build me an institutional-grade company primer on Apple (AAPL), written for a PM who doesn't follow the name closely but needs to be conversational on it in a morning meeting.
Start with the business as it actually earns money today, not the marketing version: revenue and gross profit split by product line (iPhone, Mac, iPad, Wearables and Services) and by geography, with the last three fiscal years plus the most recent reported quarter so I can see mix drift. Make the Services-versus-Products gross-margin gap explicit: segment margins, the blended rate, and how much of consolidated gross profit Services now contributes relative to its revenue share. Anchor every figure to a stated fiscal period, since Apple's September fiscal year-end makes calendar comparisons misleading.
Then give me the operating model and what actually moves the stock. Decompose the last several quarters into the drivers that explain beats and misses: units versus ASP versus mix on hardware, Services growth and take-rate structure, gross margin, tax rate, and the contribution of buybacks to EPS. Tell me which two or three disclosed metrics the market genuinely reacts to on print day and which ones it now largely ignores. Separate "will Apple beat the number?" from "will the stock go up?" Include management's guidance track record: how the range was set, where results landed, and whether forward gross-margin commentary has proved reliable.
Cover the competitive and structural position honestly. Where is the moat real (installed base, switching costs, silicon, the App Store, default-search economics) and where is it eroding or under legal attack? Size the regulatory exposure rather than merely listing it. Cover the App Store commission structure, default-search payment obligations and active litigation, with a rough earnings-at-risk estimate for the high-margin pieces being contested.
Add the China and supply-chain exposure on both sides: revenue exposure and manufacturing concentration. Flag the AI and Siri roadmap question, including what a credible outcome versus a failed outcome could mean for the device-replacement cycle.
Close with valuation and the debate. Show current and forward multiples using the peer set you believe is genuinely comparable, and explain your selection. Include a reverse-DCF assessment of the growth and margin assumptions embedded in the current share price.
Finish with a tight bull, base and bear case containing the specific swing variables, rough probability weights, and falsifiable markers that would tell me within two or three quarters which case is playing out. Name upcoming catalysts and include dates where available.
Keep the work evidence-led, with inline citations to filings, earnings calls and press releases. Where a number is not disclosed, estimate it, show the method, and state your confidence rather than skipping it. Be explicit about anything you could not verify.
That is a much longer prompt, and the length is not the point. It works because it hands the model an investment framework instead of a topic. Below we take it apart one instruction at a time.
heading: The eight techniques
Tell the AI who the work is for :: 1-tell-the-ai-who-the-work-is-for
Define the analytical lens :: 2-define-the-analytical-lens
Establish the right comparison :: 3-establish-the-right-comparison
Ask for drivers, not descriptions :: 4-ask-for-drivers-not-descriptions
Separate the company from the stock :: 5-separate-the-company-from-the-stock
Size risks instead of listing them :: 6-size-risks-instead-of-listing-them
Turn narratives into testable scenarios :: 7-turn-narratives-into-testable-scenarios
Define the evidentiary standard :: 8-define-the-evidentiary-standard
1. Tell the AI who the work is for
step: 01
Written for a PM who doesn't follow the name closely but needs to be conversational on it in a morning meeting.
One sentence about the reader changes what the whole answer is for. It tells the model that the reader knows the basics, that definitions would be a waste of space, and that the work has to be short enough to absorb before a meeting. It also says the answer should lean toward the questions likely to come up in that meeting rather than a full initiation report.
"Bring me up to speed" means something very different to a retail investor, a technology reporter, an Apple supplier and a portfolio manager. Saying who the work is for, and what they need to do with it afterward, resolves most of that in one line.
2. Define the analytical lens
step: 02
Start with the business as it actually earns money today, not the marketing version.
Without that line, the answer tends to open on history, brand, design and ecosystem. Those facts are true enough, but they are not the ones that help you. The instruction points the work at the economic engine instead: where the revenue comes from, where the gross profit comes from, how the mix is shifting, and which businesses carry the margin.
Asking for the Services-versus-Products margin gap goes one step further, because it names a relationship rather than a topic. "Analyze Services" is a subject. "Show how much of consolidated gross profit Services contributes relative to its revenue share" is a question with an answer. The second kind is what turns a summary into analysis.
3. Establish the right comparison
step: 03
The last three fiscal years plus the most recent reported quarter so I can see mix drift.
This sets the time horizon and says why it matters. Ask for Apple's revenue breakdown on its own and you will probably get the latest annual figures. Accurate, but it will not show whether the company is leaning harder on Services, how the geographic exposure is shifting, or whether the last quarter was an inflection.
The phrase "so I can see mix drift" does most of the work here: it gives the data a job. The line about Apple's September fiscal year-end does the rest, since it stops fiscal and calendar periods getting mixed together. Between them, the prompt is not just asking for data, it is fixing the basis of comparison.
4. Ask for drivers, not descriptions
step: 04
Decompose the last several quarters into the drivers that explain beats and misses.
A basic answer tells you Apple beat on revenue. A useful one tells you why. Were iPhone units stronger, or did ASPs do the work? Was the mix favorable? Did Services accelerate, did gross margin expand, and how much of the EPS line came from operations rather than the tax rate or the buyback? And the question that decides whether any of it matters: is the source of the beat likely to repeat?
A number on its own rarely tells you much. The mechanism behind it does. Asking the model to break an outcome into its drivers is usually the difference between the two.
5. Separate the company from the stock
step: 05
Separate 'will Apple beat the number?' from 'will the stock go up?'
This may be the most important line in the prompt.
A company can beat consensus and watch the stock fall anyway. The result was already priced in, or the beat came from the wrong place, or guidance disappointed, or the market had moved on to a different number entirely.
So the prompt keeps three questions apart: what drives the operating results, what investors already expect, and what would actually cause the stock to reprice. They overlap, but they are not the same question, and an answer that treats them as one reads like an earnings summary rather than an investment view.
6. Size risks instead of listing them
step: 06
Size the regulatory exposure rather than merely listing it.
Models are very good at producing risk lists. Ask about Apple and you will get App Store regulation, antitrust litigation, China, manufacturing concentration and competition, all of it accurate. The problem is that a list does not tell you which of them matters. Sizing pushes the answer further. Which revenue or profit pool is actually exposed, and what margin does it carry? What share of it could plausibly go? Does the hit land on revenue, on earnings, or only on the growth rate, and over what period would it show up? Then the one that sorts the list: is this economically material, or just prominent in the news?
Not every risk can be sized precisely, and that is fine. A rough estimate that shows its working usually beats a polished list with no sense of scale.
7. Turn narratives into testable scenarios
step: 07
Flag the AI and Siri roadmap question, including what a credible outcome versus a failed outcome could mean for the device-replacement cycle.
"Apple is behind in AI" is a narrative. It is not yet an investment framework.
The prompt asks for that narrative to be turned into something observable. What would credible progress actually look like, and could it pull forward the device-replacement cycle? What would failure look like instead? When should the first evidence either way start showing up, and which metrics would carry it?
The bull, base and bear cases work the same way. A scenario that is only optimistic or pessimistic does not help anyone. A useful one names the swing variables and says what evidence would confirm or kill it.
8. Define the evidentiary standard
step: 08
Where a number is not disclosed, estimate it, show the method, and state your confidence.
AI-generated research can read as authoritative while quietly blending reported facts, market commentary and its own inference. The prompt asks for those to be kept apart: reported facts get a citation, estimates get labeled as estimates with the calculation shown, every number carries a confidence, and anything that could not be verified is named as such rather than smoothed over.
You still have to check the work, but this makes it far quicker to check. In research, "I could not verify this" is worth more than a confident number nobody can trace.
The formula, in six parts
Every technique above collapses into six questions. The Apple prompt answers all of them, which is the only reason it works.
label: Exhibit 2 · The six-part formula
Audience :: Who will use the analysis, and what do they already know? :: users
Objective :: What should that person understand or decide afterward? :: target
Scope :: Which topics, periods and comparisons are in bounds? :: ruler
Analytical lens :: How should the problem be decomposed? :: lens
Output :: How should the answer be organized, and how long? :: list
Evidence :: Which sources, estimates, citations and caveats are required? :: evidence
None of this means every prompt has to be long. A quick definition or a simple calculation needs a sentence, and padding it out helps nobody. But as the work gets more complex, and as more rides on the answer, more of your thinking has to be said out loud rather than assumed. The aim is not more words. It is better judgment.
label: Exhibit 3 · Run this before you send
heading: Before pressing Enter, ask:
footnote: Tick them off against your own prompt. Anything you leave blank, the AI fills in for you.
Did I say who this is for?
Did I explain the decision or objective?
Did I define the relevant periods and comparisons?
Did I identify the questions that actually matter?
Did I specify the desired output?
Did I define the evidence standard?
The remaining problem: you cannot write this every time
There is an obvious catch with the Apple example: it asks the analyst to design the whole research process from scratch, every time. An experienced analyst already knows the moves:
- Decompose earnings into volume, pricing and mix
- Separate a fundamental beat from a positive stock setup
- Test management's guidance history
- Size earnings at risk
- Select economically relevant peers
- Build scenarios around observable swing variables
- Define the markers that would invalidate a thesis
A general-purpose model does not reliably know which of those steps matter until someone teaches it. And nobody wants to retype a 700-word instruction every time they pick up a familiar task. That is the gap skills fill.
From good prompts to reusable skills
Matterfact ships more than 5,000 pre-built research skills, modeled on how hedge-fund analysts actually work. Each one holds the analytical structure behind a task that comes up again and again:
- Building company primers
- Reviewing earnings
- Tracking guidance accuracy
- Sizing regulatory exposure
- Analyzing management tone
- Mapping an investment debate
- Constructing bull, base and bear cases
- Selecting the right peer set
- Monitoring the signposts that confirm or invalidate a thesis
That means the analyst states the objective rather than the method. Type a slash and the picker opens: your own saved skills on one side, the pre-built library on the other.
command: /company-primer
argument: Apple
note: Type / and the picker opens. Your own saved skills on the left, the pre-built library on the right.
cross-company-read-through :: Quantify read-through from one print to its peers
company-primer :: Institutional-grade company overview
guidance-track-record :: Score management against its own guidance
regulatory-earnings-at-risk :: Size the exposure, not just list it
Matterfact identifies the relevant skills, performs the underlying research steps, and produces a structured output that can be customized to the analyst's or firm's process.
label: Exhibit 4 · From request to research
Analyst request :: One line. "Build me a PM-ready primer on Apple."
Skills selected :: Business model, KPI decomposition, guidance history, debate mapping, risk sizing, valuation
Research performed :: Filings, earnings calls, market data, internal research and firm-specific sources
Output produced :: A cited, institutional-grade primer in the firm's own format
The takeaway
Learning to prompt well is still worth the effort. It makes you state your objective, your assumptions and your judgment out loud, and that is useful well beyond the chat box. It also makes a bad answer easier to diagnose: was the data wrong, was the reasoning weak, or did the prompt never ask the real question in the first place?
What it should not mean is rebuilding the same workflow from scratch every week. The frameworks worth keeping end up as reusable skills, improved over time and shared across the firm.
A good prompt transfers your judgment to the AI.
A good skill makes that judgment repeatable.
Next weekend
In the next edition of the Matterfact Weekend Education Series, we will look at how to evaluate an AI-generated research report, and how to catch confident-sounding answers that are incomplete, unsupported or simply wrong.
FAQ
What makes a good AI prompt for investment research?
A good research prompt transfers your analytical intent, not just your topic. It names the audience, the decision the work supports, the periods and comparisons that matter, how the problem should be decomposed, the shape of the output, and the evidence standard the answer must meet. Length is a symptom rather than the goal: what matters is whether your intent made it across.
Are longer prompts better?
No. Longer prompts are only better when the extra words carry analytical judgment the AI would otherwise have to guess at. Simple tasks deserve simple prompts. Complexity in the prompt should scale with the complexity and importance of the work.
Why does AI give generic answers to investment questions?
Because a vague question leaves every analytical choice to the model, and models default to the safest possible answer. Ask "is Apple a good stock to buy?" and the AI has to guess your time horizon, your benchmark, which drivers you care about, and what evidence you would accept. It guesses conservatively, which produces a broad, competent and forgettable summary.
How do I stop AI from producing a risk list instead of risk analysis?
Ask it to size the risk rather than list it. Specify which revenue or profit pool is exposed, what margin that pool carries, what portion could plausibly be lost, whether the impact hits revenue, earnings or only the growth rate, and over what period. A rough, transparent estimate beats a polished but unprioritized list.
What is the difference between a prompt and an AI skill?
A prompt transfers your judgment to the AI once. A skill captures that judgment as a reusable workflow so you do not rewrite it every time. Matterfact ships more than 5,000 pre-built research skills modeled on hedge-fund analyst workflows, so an analyst can ask for a PM-ready primer in one line instead of specifying every analytical step by hand.