Asking "is it worth buying shares of X" is the worst possible prompt: the model will respond politely, generally, and from memory a year ago. This guide shows a different way to work with AI when analyzing a public company – a sequence of seven prompts in a fixed order, in one conversation, with strict rules for citing numbers and a checklist of things to verify before trusting the output. It works in Claude, ChatGPT, and Gemini, for companies from the USA and the Warsaw Stock Exchange.
Last verified: September 6, 2026. Informational and educational material – not investment advice or a recommendation. AI tool functionalities change every few weeks; confirm details in the manufacturer's documentation. Investing in stocks involves the risk of losing all invested funds.
In Brief
- Without web search enabled, the model relies on memory, which ends at the training date. In Claude, you enable search from the “+” menu under the text box (Anthropic instructions), in ChatGPT the search mode works (OpenAI help), in Gemini the Deep Research mode (Google description).
- Order matters: first context and facts, then industry background, and only then attack the thesis, scenarios, and notes for review. A model that hears your opinion at the start will defend it instead of verifying it.
- One conversation per company. Each prompt builds on the previous responses. A new chat for each step disrupts the entire scheme.
- A number without a source and date does not exist. Each prompt contains the same rule: company documents first, secondary services later, and anything the model cannot find should be stated as "I did not find it," not estimated.
- Two numbers verified in the report from the investor relations page or in the EDGAR database take five minutes and indicate how much the rest of the answers are worth.
- The result has an expiration date. Prices and multiples change daily, and reports are released quarterly. Therefore, the last prompt turns the analysis into a note with threshold values, which you return to after each report.
Why a Simple Question About a Company Doesn't Work
Language models have three habits that cost money when it comes to stocks. First, they respond from memory: if you don't instruct them to search, they will provide revenue from four quarters ago as current and will not indicate that it is outdated. Second, they confirm what they have heard: if you start with "I think this is a great company," you will receive a list of arguments in favor, while risks will be smoothed over with a single sentence, "it is worth remembering." Third, they fill in the gaps: they can calculate a missing operating margin "approximately" and present it without reservations as if it came from the report.
The American stock market regulator additionally warns that the label "AI" today is a tool for fraudsters: promises of "an algorithm that picks winning stocks" are a classic red flag (SEC warning on investor.gov). This guide goes in the opposite direction: AI should not predict anything. It should gather facts, organize them into tables, play against you, and record at what numbers you would change your mind. The decision remains yours.
The right metaphor – treat the answer like a note from a capable intern with access to a library: they gather quickly, organize well, but sometimes get things wrong. You check a sample, not believing the whole.
Before You Paste the First Prompt
Choose a Tool and Enable Web Access
Enable search from the “+” menu under the text box; with search enabled, the model can also retrieve the content of a specific link you paste, e.g., a quarterly report (documentation). It handles long, multi-point instructions well.
Web search is available in both free and paid plans (OpenAI help). For the first, longest step, you can use the Deep Research mode, which browses multiple sources and returns a report with citations (feature description).
The Deep Research mode builds a search plan, browses pages, and returns a report with sources (Google description). It is suitable for step 1; subsequent prompts are pasted into a regular conversation in the same window.
All three can execute this scheme. The differences mainly concern how long the model adheres to the rules from the first prompt: when it starts to "forget" the citation rule, remind it with a single sentence instead of starting over.
One Calibration Question
Before you start the analysis, check if the model is indeed searching the web and understands the company's calendar. Paste:
Before we begin: provide today's date, the name and publication date of the company's latest quarterly report [COMPANY], the end date of its fiscal year, and the currency in which it reports. Add a link to the source for each answer.
If the dates are from six months ago or lack a link, the model is not searching. Respond: "Search the web and correct before we proceed." The end of the fiscal year is more important than it seems: Nvidia has a fiscal year ending on the last Sunday of January (10-K annual report in the SEC database), so its "fourth quarter of 2027" is practically November 2026 – January 2027. A model that does not know this difference will mix up quarters in the table, and no one will notice.
Where the Model Should Search
In the prompts, there is a rule of "company documents first." It is worth knowing what this specifically means, as you can provide the model with an address or even paste a link to a specific report.
| What You Need | US Companies | GPW Companies |
|---|---|---|
| Annual and quarterly reports | EDGAR full-text search (forms 10-K, 10-Q, 8-K; for foreign companies 20-F and 6-K), the company's investor relations page | Company's investor relations tab, ESPI and EBI announcements on the GPW website |
| What is 10-K and why read it | SEC explanation for investors | Periodic report according to IFRS; published in the ESPI system |
| Insider transactions | Forms 3, 4, and 5 (SEC explanation) | Notifications of transactions by management in ESPI announcements |
| Short selling | FINRA short position data | Short selling register from KNF |
| Quotes and ratios | Investor relations page, financial services (secondary data) | Stooq, Bankier (secondary data) |
| Risk-free rate for comparisons | US Treasury yield curve, DGS3 series in the FRED database | Retail bond interest rates |
Ticker in One Place
Each prompt is in a separate box with a header and a “Copy prompt” button below. You copy the entire box, paste it into the chat, and replace the brackets. In the prompts below, the brackets are: [COMPANY], [COMPETITORS], [AMOUNT], [CURRENCY]. You replace them once, in the first line, and the rest works unchanged. For comparison, take three to six companies from the same industry and of similar size; comparing a bank with a game producer gives a nice table and zero information.
Scheme: Seven Steps in One Conversation
- Investor Context
You tell the model what currency you are counting in, for how long, what portion of the portfolio you are considering, and what you already have. Without this, the opportunity cost and position size will be calculated for someone else.
- Company Analysis
Fact base: business model, eight quarters in a table, advantages, threats, balance sheet, capital allocation, insiders. It ends with a briefing that moves on.
- Industry Background
Three tables against the competition, all numbers from the same day and on the same basis. Answers the question of how much growth and quality you get for a unit of valuation.
- Pre-mortem
The model assumes you lost and looks for reasons in today's data. No softening. It ends with a condition after which the black thesis ceases to apply.
- Three Scenarios
Value ranges and, more importantly, the answer to the question: what assumptions need to be made for today's price to be fair.
- Quarterly Note
One page with a thesis, three numbers, three events, and thresholds. This is the only element you enter into the calendar.
- Post-Report Review
The only step in a new conversation: you paste the note, the model retrieves the new report, and checks threshold by threshold. Without this, the analysis ages in the chat.
Why in this order. Facts before opinion, because the model anchors on the first thing it hears. Industry background before attack, because pre-mortem without comparison to competitors ends up with generalities (“competition may intensify”). Scenarios after pre-mortem, because only then does the pessimistic scenario have concrete content. The note at the end, because it summarizes everything that was above.
Prompt 0 – Investor Context
This step is short but changes the outcome of the two later ones. In step 4, the model calculates the opportunity cost: for someone counting in zlotys, the reference point is treasury bonds in PLN, for someone counting in dollars, US bills and bonds. For the Polish diaspora, there is also the issue of which country you settle capital gains tax; we describe residency rules in a separate guide on tax residency and PIT.
We are starting the analysis of the company [COMPANY]. Before you calculate anything, remember my context and apply it in all subsequent answers in this conversation:
– I am counting in currency: [CURRENCY]. Convert all amounts and opportunity costs to it, providing the exchange rate and date.
– Horizon: [e.g., 3 years]. I am not interested in price movements on a weekly scale.
– Considered position size: [e.g., 5 percent of the portfolio, about AMOUNT].
– Rest of the portfolio: [e.g., broad US stock index, treasury bonds, two tech companies]. Pay attention to whether this company adds risk that I already have.
– I settle capital gains tax in: [country]. Do not advise on taxes, but when calculating net profit, indicate that it is before tax.
Confirm in three sentences how you understand this context, and do not evaluate the company yet.
Check in the response: whether the model has already started praising or criticizing the company. If so, reply: “Do not evaluate yet. Wait for the next prompt.” The goal is for the first thing it says about the company to be facts from the search, not a reflex.
Prompt 1 – Company Analysis
The longest step. It builds the fact base on which all subsequent ones stand. Eight quarters instead of four, because four quarters do not show margin trends or seasonality. It includes things that financial service tables are silent about: customer concentration, geographical exposure, stock dilution, the accuracy history of management forecasts, and what insiders are doing with their shares.
You are an equity analyst preparing material for an individual investor. Conduct a full analysis of the company [COMPANY].
DATA RULES
– Use search. Every number has a source: document name or service, publication date, and link.
– Source order: first company reports (quarterly, annual, earnings presentation, analyst call transcript), then regulator data, finally secondary services. Indicate with a letter where each number comes from: [S] company, [R] regulator, [W] secondary service.
– What you cannot find, do not estimate. Write “I did not find” and indicate in which document to check.
– Provide reported data (GAAP or IFRS). If the company highlights a “non-GAAP” version, show both and the difference between them.
– At the top of the response: today's date, date of the last quarterly report, end of the fiscal year, reporting currency.
DELIVER IN THIS ORDER
1. Where the money comes from. One paragraph on the business model, followed by a breakdown of revenue by segments in percentages and by geographical regions in percentages for the last reported quarter.
2. Table of the last eight quarters. Columns: quarter, revenue, year-over-year change, gross margin, operating margin, diluted earnings per share, free cash flow, number of diluted shares. Below the table, one sentence about seasonality if it is visible.
3. Concentration. What percentage of revenue does the largest customer provide and the top three combined. What percentage of costs depends on one supplier or one country. If the company does not disclose this, state it directly.
4. Competitive advantage. Name the mechanism (switching costs, network effect, scale, patents, regulation, long-term contracts, brand) and add one numerical piece of evidence that confirms or undermines this mechanism.
5. Who threatens this advantage in the next 24 months. Three names, with one sentence each: how specifically and from when.
6. Management credibility. Compare the last four quarterly forecasts from management with the actual results. How many times did they hit, underestimate, and overestimate.
7. What insiders and the market are doing. Transactions by management in the last six months (buy or sell, value, whether it is a pre-established sales plan). Percentage of shares sold short and how it has changed in three months.
8. Balance sheet in four numbers: cash and liquid investments, interest-bearing debt, net debt to EBITDA, debt maturity dates in the next three years.
9. Capital allocation over the last twelve months: stock buybacks, dividends, acquisitions, capital expenditures, stock-based compensation. Separately: does stock buyback exceed dilution from incentive programs.
10. Catalysts for the next 12 months: date or time window, event, direction of impact, and your assessment of whether the market has already priced this in (with an argument).
11. Uncertainty card: list of numbers from this response that you could not confirm in the company's document.
12. BRIEFING FOR STEP 2: six sentences, just facts with numbers, zero evaluations. We will return to this.
FORMAT
Short paragraphs and tables. No introduction, no summary, no recommendations. Do not say whether to buy. Separate facts from their interpretation: start interpretation with the word “Interpretation:”.
Check in the response:
- Four header items: today's date, report date, end of the fiscal year, currency. The absence of any means the model took shortcuts.
- Labels [S], [R], [W]. If most numbers have [W], ask to replace them with numbers from the reports: “Replace numbers from secondary services with numbers from the quarterly report; if they differ, show both.”
- Take two numbers from the table, revenue and diluted earnings per share from the last quarter, and compare with the report on the investor relations page or in the EDGAR. If both match, you can trust the rest more. If one does not, check the entire table.
- Point 6 (management forecasts) often returns empty. This is important information: either the company does not provide forecasts, or the model did not find them. Ask which of the two.
- Point 11 should not be empty. A model that “confirmed everything” did not check.
Prompt 2 – Industry Background
The second step answers two questions: how much growth and quality you get for a unit of valuation, and whether a cheap company is cheap for a reason you agree with. It adds a third table with three-year trends, as a snapshot from one day does not distinguish a company that is accelerating from one that is slowing down.
The same conversation. Set [COMPANY] next to competitors: [COMPETITORS].
DATA RULES
– All numbers from the same day. Provide this date above the first table.
– One basis for all: data for the last twelve months, not forecasts. Write this under each table.
– One currency: [CURRENCY]. For companies reporting in a different currency, provide the exchange rate and date of conversion.
– A loss-making company has no price-to-earnings ratio. Enter a dash, not a number. The same for negative EBITDA.
– Under each table, sources column by column.
TABLE 1, valuation vs. growth
Rows are companies. Columns: market capitalization, enterprise value, year-over-year revenue growth, gross margin, operating margin, price-to-sales, price-to-earnings, enterprise value to EBITDA, free cash flow yield (free cash flow divided by market capitalization), and your own column: revenue growth divided by price-to-sales.
TABLE 2, quality of business
Columns: return on equity, return on invested capital, free cash flow margin, net debt to EBITDA, year-over-year change in the number of shares, stock-based compensation as a percentage of revenue.
TABLE 3, three-year trend
Columns: average annual revenue growth over three years, change in operating margin in percentage points over three years, change in the number of shares over three years, average price-to-sales from three years next to today’s.
RANKING
Rank companies from best to worst based on how much growth and quality I get for a unit of valuation. Label each: LEADING, AVERAGE, or LAGGING, and justify with one sentence with a specific number from the table. Separately, write how the ranking would change if forecasts were taken instead of historical data, and why we do not do that by default.
PITFALL OF CHEAP VALUATION
Identify the company with the lowest multiples and determine: cheap because the market undervalues it, or cheap because the business is shrinking. Three numbers that decide this, from Table 3.
WHAT THESE TABLES DO NOT SHOW
Three things outside of multiples that could reverse this ranking within a year (e.g., regulatory change, one contract, inventory cycle, currency exchange rate). For each, write where to check this in a document or service.
Check in the response:
- Whether the price-to-earnings ratio has the same basis for all. The model likes to take historical data for one company and analyst forecasts for another; then the ranking means nothing. Ask directly: “For each company, write what period the earnings in the denominator come from.”
- Whether there is a dash next to the loss-making company. A number in this place means the model took the metric from a service that calculates it differently.
- Whether market capitalization and enterprise value differ. If they are identical for a company with significant debt, something is wrong.
- The “stock-based compensation” column is often omitted. For tech companies, this is often several to dozens of percent of revenue, explaining why “stock buybacks” do not reduce their number.
Prompt 3 – Pre-mortem
The pre-mortem technique comes from project management: instead of asking “what could go wrong,” you assume the project has already failed and look for reasons (Gary Klein's article in Harvard Business Review). In stock analysis, it works better than a regular risk list because it forces specificity: not “competition may intensify,” but “in which quarter and in which number is this visible.” The most important point is the last one: the condition after which the black thesis ceases to apply. This is the only thing from this step that goes into the calendar.
The same conversation. Now you are playing against me. Assume I bought [COMPANY] at today’s price and in two years I am at a loss. Write why.
RULES
– Do not soften or balance. Not a single sentence starting with “on the other hand,” “it is worth remembering,” “one should keep in mind.”
– Each accusation based on a number or quote from the company document, with a link. Risks without evidence should be removed from the list.
– No catastrophes. I am interested in what is already visible today in the data from steps 1 and 2.
DELIVER
1. Three red flags. For each: weight (high, medium, low), evidence with a number and source, mechanism, i.e., through which position in the income statement or balance sheet this thing turns into a drop in price, and the quarter in which it will be visible at the earliest.
2. Accounting check. Check and provide results for each point: do receivables grow faster than revenue; do inventories grow faster than revenue; does net income grow while operating cash flow does not; how large is the difference between reported profit and “non-GAAP” and where it comes from; does the company capitalize costs that competitors expense immediately. For each: number, source, assessment.
3. Determine whether this is a company at the bottom of the cycle before recovery or a value trap. Three numbers from the last eight quarters that decide this.
4. Scenario “the price goes sideways for three years.” What would have to happen to revenue, margin, and multiple. Numeric ranges.
5. Hidden cost. How much do I lose by holding [AMOUNT] in this company instead of in treasury bonds in my currency or in a broad index if the price stays flat for three years. Provide the current interest rate with a source and date. Indicate that the result is before tax.
6. One condition after which this entire black thesis ceases to apply. It should be verifiable in a specific document or on a specific date so that I can enter it into the calendar.
7. Three questions you do not know the answers to that would change this picture. For each: where to look for answers.
Check in the response:
- Soothing sentences despite the ban. Models default to smoothing criticism, and this is the biggest loss of value in this step. Reply: “Remove all soothing sentences, leave only accusations.”
- Point 6. If the condition reads “improvement in sentiment” or “return to growth,” that is not a condition. A condition has the name of a position in the report, a threshold value, and a date.
- Point 2 sometimes returns as “no data.” Receivables and inventories are in every balance sheet; if the model did not find them, ask it to retrieve the latest quarterly report and calculate directly.
- Currency check in point 5: for someone counting in zlotys, the comparison is treasury bonds in PLN, for someone counting in dollars, US bonds. A model that does not remember the context from prompt 0 will compare what it should not.
Prompt 4 – Three Scenarios and What Today's Price Assumes
This step is missing in most prompt guides, yet it closely answers the question of buying and selling. Instead of asking “how much is this company worth” (the model will give one number, which will be worthless), you reverse the task: what assumptions about growth, margin, and multiple need to be made for today’s price to be fair? If these assumptions are more optimistic than anything the company has delivered in the last eight quarters, you know what you are buying.
The same conversation. Build three scenarios over a three-year horizon for [COMPANY]: pessimistic, base, and optimistic.
RULES
– Each scenario has four assumptions: average annual revenue growth, operating margin in the third year, number of shares in the third year, valuation multiple at the end (price-to-earnings or enterprise value to EBITDA, the same for all scenarios in that row). Justify each assumption with one number from steps 1 to 3: the company’s history or that of its competitors.
– Do not use market averages without stating where they come from.
– Do not provide a “target price.” Provide ranges and assumptions.
DELIVER
1. Scenario table: rows are scenarios, columns are four assumptions, resulting value per share, and annual return from today’s price. Below the table, assign a probability to each scenario and write what you base it on.
2. Reverse question: what assumptions need to be made for today’s price to be fair at an expected return of [e.g., 10 percent per year]? Compare these assumptions with what the company actually delivered in the last three years. One sentence: does the market assume acceleration, maintenance, or deceleration.
3. Sensitivity: which of the four variables most affects the result? Show how much the value per share changes when you shift that one variable by one percentage point or one multiple point.
4. Asymmetry: the ratio of profit in the optimistic scenario to loss in the pessimistic one, counting from today’s price. One sentence of interpretation, without recommendations.
Check in the response: whether the multiple at the end of the scenario is justified by something more than “historical average.” The model likes to take today’s often high multiple and assume it will hold for three years. Ask: “What multiple did this company and its competitors have three years ago and five years ago?” Second check: number of shares. If it stays the same in the base scenario, while step 2 showed it growing by two percent annually, the table overstates the value per share.
Prompt 5 – Quarterly Note
Analysis without an expiration date turns into a hunch. This prompt turns five long responses into one note, which you return to with the next report. The threshold values recorded today tell you directly when to change your mind, and they do so at a time when you are already emotionally attached to the position.
Wrap these steps into one note that I will return to after the next quarterly report.
– Thesis in three sentences: what this company must deliver for me to profit.
– Three numbers that confirm this thesis: today’s value, reading date, source.
– Three events that break it: each with a date or document in which I will check this.
– Thresholds for the next report: name of the position in the report and value below which I consider the thesis broken. Separately, the value above which the thesis strengthens.
– Condition from the pre-mortem after which the black thesis ceases to apply.
– Date of the next quarterly report and the date when I should return to this note.
– One sentence: what I still do not know.
No introduction, no summary. Just points, in a format I can paste into notes.
Check in the response: whether the thresholds are numbers, not adjectives. “Operating margin below 28 percent” is a threshold. “Significant deterioration of margin” is not. Save the note outside the chat; after a quarter, the conversation may be unavailable or too long to retrieve.
Prompt 6 – Post-Report Review
This is the only step you take in a new conversation, usually a few days after the report is published. You paste the note from step 5 and instruct the model to check threshold by threshold. You do not ask “what do you think of the results,” because you will get a summary of the company’s press release. You ask whether the numbers exceeded the thresholds you set when you did not yet have a position.
Web search enabled. Below, I paste my note with the thesis, numbers, and thresholds for the company [COMPANY], written on [NOTE DATE]. The company has published a new quarterly report.
RULES
– Retrieve the quarterly report and earnings presentation from the company’s website or from the regulator’s database. Provide dates and links.
– Reported data, not “non-GAAP,” unless the threshold in the note was defined differently.
– Do not evaluate the results. Only assess whether they exceeded my thresholds.
DELIVER
1. Table: each threshold from the note, threshold value, value from the new report, verdict (maintained, broken, no data), source.
2. Three numbers confirming the thesis: previous value, new value, direction.
3. Events breaking the thesis: did any occur since the note date? Look for this in company announcements and in the analyst call transcript.
4. What changed in the report that was not on my note: new risk, new segment, change in management forecast, change in management. Only from documents, with a link.
5. Updated note in the same format, with new values and date.
NOTE:
[paste note from step 5]
Check in the response: whether the verdicts in point 1 result from numbers, not from interpretation. If the threshold read “revenue grows above 15 percent year-over-year,” and the model writes “maintained because management expects acceleration,” that is broken. Management’s forecast is not a result.
Appendix: One Prompt for the Entire Portfolio
After running several companies through this scheme, you will notice that their notes have common risks. It is worth asking about this directly once a quarter, in a new conversation:
I am pasting notes for the companies I have in my portfolio, with weights: [list: company, percentage of portfolio]. Find risks that repeat in at least two notes (the same end customer, the same investment cycle, the same regulation, the same currency, the same commodity). For each repeating risk, provide the total percentage of the portfolio that depends on it, and one event that would trigger it in all companies at once. No recommendations, just a map of dependencies.
How to Check the Response in Five Minutes
You do not need to verify everything. A sample is enough, but a well-chosen one.
- Dates at the top. A report date older than the last quarter means the model did not search or hit an old article.
- Two numbers from the table. Revenue and diluted earnings per share from the last quarter. In the EDGAR full-text search, type the company name and the phrase “total revenue,” narrowing it to form 10-Q; for a GPW company, open the periodic report in the investor relations tab. Compare.
- One number from the “secondary data” column. If the model marked it [W], check if it differs from the report. Financial services often calculate margins and ratios differently.
- One link. Click on a random source. Sometimes the link exists but does not contain the cited number. This is a signal to check the rest.
- Basis of price-to-earnings ratio. Historical data or forecast. One question to the model is enough.
Common Numerical Pitfalls AI Often Falls Into
| Pitfall | How It Appears in the Response | How to Catch It |
|---|---|---|
| Historical data mixed with forecasts | Price-to-earnings of 25 for one company (from the last 12 months) and 18 for another (analyst forecast) | Ask to provide the period of earnings in the denominator for each company |
| Reported profit vs. “non-GAAP” | High margin that disappears after adding stock-based compensation and write-offs | Require both versions and the difference between them |
| Fiscal year different from calendar year | “Q4 2026” for a company with a year ending in January or September | Calibration question about the end of the fiscal year |
| Stock split | Earnings per share before the split compared with earnings after the split; “drop” of 90 percent | Ask whether the data is adjusted for stock splits |
| Currency and depositary receipts | Company reports in euros, receipts traded in dollars; multiples calculated on a mix | One currency, exchange rate, and date of conversion under the table |
| Basic vs. diluted earnings per share | Higher earnings per share for a company with a large options program | In prompts, request diluted |
| Revenue vs. orders | “Revenue increased by 60 percent,” when the order book grew, and revenue by 12 | Ask about the position in the income statement, not from the presentation |
| Negative ratios | Price-to-earnings “minus 40” or enterprise value to EBITDA for negative EBITDA | Rule “dash instead of number” in prompt 2 |
| Stock buybacks that buy nothing | Billions for buybacks, while the number of shares increases because incentive programs give more | Column “year-over-year change in the number of shares” |
Ratios Depend on the Industry
The tables in prompts 1 and 2 are written for a typical manufacturing, technology, or service company. For several industries, standard multiples are misleading and need to be replaced in the prompt. Just add in the rules: “In Table 1, replace columns X and Y with columns Z.”
| Industry | What Doesn’t Work | What to Insert Instead |
|---|---|---|
| Banks and insurers | Enterprise value, EBITDA, free cash flow (debt is a commodity here, not a problem) | Price to book value, return on equity, capital ratio, cost of risk, share of non-performing loans |
| Real estate (REIT) | Net income (understated by depreciation) | Operating cash flow per share, debt to asset value, occupancy, average lease duration |
| Subscription software | Price to earnings (often no earnings) | Growth of recurring revenue, net revenue retention, gross margin, growth plus free cash flow margin (the so-called rule of 40) |
| Cyclical companies (chemicals, steel, automotive, semiconductors) | Multiples at the peak of the cycle look cheap, at the bottom expensive | Multiple to average earnings over the full cycle, net debt, capacity utilization |
| Retail | Revenue growth alone (may come from new stores) | Comparable sales in existing stores, gross margin, inventory to sales, revenue per square meter |
| Commodities and energy | Historical margins at a different commodity price | Cost of extraction per unit, breakeven threshold at commodity price, reserve life, replacement costs |
| Biotechnology without revenue | All multiples | Cash and months until it runs out, research stage, regulatory decision dates, dilution from issuance |
Version for GPW Companies
The scheme works the same way, but sources and a few rules require an addition. To the data rules in prompts 1 and 2, add:
The company is listed on the GPW. Priority is given to periodic reports from the company (IFRS report and management report) and ESPI announcements published on the GPW website. Take quotes and ratios from Stooq or Bankier services and mark them as secondary data. All amounts in zlotys. Also provide liquidity: average daily turnover from the last three months and the percentage of shares in free float.
Three differences to keep in mind:
- Current ESPI reports are the main channel for price-sensitive information in Poland: contracts, insider transactions, changes in management, forecasts, and their corrections. A model to which you point the GPW announcements page will find more there than in press articles.
- Liquidity and free float. For many GPW companies, the biggest risk is not the business but that you cannot sell a package without moving the price. Hence the addition about turnover.
- Tax. A Polish resident settles gains from stocks and dividends in PIT-38 at a rate of 19 percent; from 2027, a Personal Investment Account will be introduced, which we describe in a separate guide on OKI and Belka tax. A US resident with Polish stocks settles them in the USA; with Polish funds and ETFs, an additional PFIC regime appears, which we discuss in the guide on saving in Poland from abroad. The model will not replace an advisor here but can indicate that the result is before tax.
Security and Privacy
- Brokerage account statements with account numbers, personal data, or full account balances. In prompt 0, an approximate amount and percentage of the portfolio are sufficient.
- Login data for banks or brokers, even if the assistant promises to “check the balance.”
- Confidential information about the company if you have access to it due to work; using it for investing is a crime regardless of whether it is processed by a human or a model.
Also check training settings: in Claude, the decision to use conversations for model training depends on account privacy settings (Anthropic article), in ChatGPT similar switches are in the “Data controls” section (OpenAI help). If someone offers you an “AI bot” that will invest for you for a fee, check our guides on guaranteed returns and investment groups on WhatsApp; the scheme is the same, only the word “AI” is new.
Common Mistakes
- Starting from your own thesis. “I believe X is undervalued, confirm” gives confirmation. Always.
- New chat for each prompt. Step 3 without steps 1 and 2 is a list of generalities about risk.
- Search disabled. The answer looks identical, only the numbers are from a year ago.
- Treating a “7 out of 10” rating as a recommendation. This is a number invented by the model based on the text it wrote itself.
- Comparing companies from different industries in one table. Multiples of a bank and a software company do not mean the same thing.
- Saving the note in the chat. After a quarter, you will not be able to retrieve it. Notes, calendar, file.
- Omitting step 6. A one-time analysis checks whether to buy. Only the post-report review checks whether to hold.
- Believing in the link. The presence of a link does not mean that the page contains the cited number. Click on one.
This scheme organizes research and forces a look at the other side. It does not predict prices and does not tell you what to buy. The model may provide a number that does not exist, and it will do so with certainty. The decision and risk are yours; you can lose all the money invested in stocks. For larger amounts, consult a licensed advisor.
Frequently Asked Questions
Which model is best for stock analysis?
The one where you have web search enabled and whose response you have checked against two numbers. The differences between Claude, ChatGPT, and Gemini are smaller than the difference between a response from the web and a response from memory. If you have access to deep research mode, use it for step 1; the remaining steps require conversation, not a report.
Can I paste the quarterly report to the model instead of asking it to search?
Yes, and this often yields better results: with search enabled, Claude can retrieve the content of the link, and ChatGPT and Gemini can attach a PDF file. Then add in the rules: “Take numbers first from the attached document.”
How long does the entire scheme take?
Step 1 takes a few minutes of model work, steps 2 to 5 about a minute or two each. With reading and checking two numbers, count about an hour per company. Step 6, quarterly, about fifteen minutes.
What if the model refuses to provide a rating or scenarios?
Some models treat questions about stocks cautiously. It helps to remind them that you are asking for an analysis of facts and scenarios with assumptions, not a recommendation, and that you make the decision yourself. All prompts in this guide are formulated this way.
Does this work for ETFs and funds?
Steps 1 and 3 only partially. For an ETF, costs, index replication, liquidity, and legal structure are more important; for US residents, also the PFIC status of funds outside the USA. This is a separate topic.
How often should I repeat the ranking from step 2?
The ranking is current on the day you made it. Multiples change with the price. In the post-report review (step 6), you can add: “Refresh Table 1 from step 2 with today’s data,” but then in the same conversation as the original analysis or with pasted tables.
Fact-Check Summary
- Definitely true: Claude, ChatGPT, and Gemini have web search modes, described in the manufacturers' documentation (links in the text). Company reports from the USA are publicly available in the EDGAR database, and those from GPW in the ESPI system. Nvidia has a fiscal year ending on the last Sunday of January, as confirmed by its 10-K report.
- Probably true: language models more often confirm the user's thesis and smooth criticism than seek it; this is an observation from practice, consistent with why manufacturers recommend precise instructions in prompts.
- What is uncertain: the quality of search and fidelity of citations change with each version of the model. Prompts from this guide worked in September 2026; in a few months, they may require minor adjustments.
- Common myth: “AI will predict the price.” It will not predict. It can gather data, organize tables, and find holes in the thesis. A rating of “7 out of 10” is an interpretation of text, not a forecast.
Comments (0)
No comments yet. Be the first!