Financial statements are the key to understanding a company, but the three statements pack hundreds of line items and the critical numbers hide in footnotes; valuation by gut feel produces "target prices" that don't survive scrutiny; risk analysis degenerates into boilerplate like "competitive pressures intensifying." Toss a raw "analyze this company" or "value this stock" at an LLM and it either fabricates a neat set of market data that doesn't exist, or parrots management's framing as if it were ground truth. This pack is split into three levels: beginner extracts key line items from the three statements and flags anomalies, intermediate builds a DCF and multiples framework with peer benchmarking, expert runs four-dimension risk decomposition with bull/bear logic and sensitivity. Each prompt anchors on the source text, declares that all data must be manually verified, and ends with a not-investment-advice disclaimer--three things that pull the AI from "fabricating data as an analyst" back to "assistance tool." Aimed at investors, analysts, finance professionals--anyone who reads financial reports and doesn't want to be led astray by AI-fabricated numbers.
Recommended models: financial reports often run over a hundred pages. For long-document parsing and numerical reasoning, Claude (long context, stable on tables and footnotes) or DeepSeek (long reasoning chains, suited to valuation calculations and logical deduction) are recommended.
Beginner: Quick-Read the Three Financial Statements
When you have a financial report (annual or quarterly) and want to quickly pull the key line items from the three statements and spot anomalies. The point isn't getting the AI to "write an analysis report"--it's getting it to extract numbers item by item and flag what's off: revenue surging but cash flow negative, receivables growing far faster than revenue, inventory turnover slowing.
You are a senior financial analyst. Based on the financial report I provide, output a quick-read of the three statements:
1. Income statement key items (latest period + YoY):
- Revenue, COGS, gross profit / gross margin
- Selling expenses, G&A expenses, R&D expenses
- Operating profit, net income
2. Balance sheet key items:
- Cash & equivalents, accounts receivable, inventory
- Total assets, total liabilities, debt-to-asset ratio
- Net assets (shareholders' equity)
3. Cash flow statement key items:
- Operating cash flow (net)
- Investing cash flow (net)
- Financing cash flow (net)
- Free cash flow (operating CF - capex)
4. Anomaly flags: list 3-5 items warranting attention, each tagged with a type:
- Revenue vs. cash flow divergence (profit up but operating CF negative)
- Receivables / inventory abnormal growth (growth rate significantly exceeds revenue)
- Margin shift (gross / net margin swinging sharply)
- Debt structure change (short-term borrowings spiking)
- Non-recurring items dominating (large gap between net income and adjusted net income)
Constraints:
1. Extract only numbers that genuinely exist in the report I provide; do not fabricate or fill in missing data
2. For missing data, mark "not disclosed in source" rather than guessing
3. All data must be manually verified against authoritative sources (exchange filings / audit reports)--the AI may misread or skip rows
4. Avoid empty evaluative language; every conclusion must be backed by a specific number or fact
5. End with: "The above is AI-assisted analysis, not investment advice, and must be reviewed by a qualified professional"
Financial report text / data: {{paste the report or key sections}}Why these constraints: financial data is factual information, and the AI's most common error is "inventing a plausible but nonexistent number"--if you only paste the income statement, it may fill in cash flow figures too. Forcing "extract only what exists" and "mark missing as not disclosed" plugs that hole. Anomaly flags use type tags to pull the AI from "vaguely noting risk" to "which specific line item is off"--e.g., revenue up 30% but operating cash flow negative, the AI should tag it "revenue vs. cash flow divergence" and cite the specific numbers. Free cash flow gets its own line because it reflects the company's actual cash-generating ability better than net income.
Intermediate: Company Valuation & Peer Comparison
When you've read the three statements and want to go further: build a valuation, pull peer comps. The point isn't getting the AI to hand you a "target price"--it's getting it to lay out the DCF and multiples framework, list every assumption, and pull comparable companies. The actual valuation judgment stays with you.
You are a senior valuation analyst. Based on the company financials and industry information I provide, build a valuation framework:
1. Multiples (relative valuation):
- Select 3-5 peer companies; list their P/E, P/B, EV/EBITDA, P/S multiples
- Calculate the target company's current multiples; compare to peer median / mean
- Flag whether multiples are high or low; suggest possible reasons (growth differential, risk premium, cycle position)
2. DCF (absolute valuation):
- List 5 key assumptions for free cash flow forecasting: revenue growth, gross margin, capex ratio, tax rate, WACC
- For each assumption, give bullish / base / bearish cases with stated rationale
- Output the valuation range for each case; do NOT give a single "target price"
3. Assumption checklist: a table listing every assumption, its value, and its source (industry reports / historical data / management guidance)
4. Cross-check: do the multiples and DCF results agree? Where do they diverge, and why?
Constraints:
1. Peer multiples must be tagged with a data timestamp; note "must be manually verified against latest market data--AI may give stale or fabricated multiples"
2. Do not fabricate stock prices, market caps, or latest financial data; for anything not provided, mark "to be supplemented"
3. Every assumption must be traceable; the AI may not fill in numbers without an explained source
4. Clearly distinguish "AI-computed valuation range" from "investment decision"--valuation is a tool, not a conclusion
5. Avoid empty evaluative language; every conclusion must be backed by a specific number or fact
6. End with: "The above is AI-assisted analysis, not investment advice, and must be reviewed by a qualified professional"
Company financials + industry: {{paste company financials key data + industry overview}}Why this structure: the biggest trap in valuation is a "black box target price"--the AI gives you a number but you don't know where the assumptions came from. Requiring each assumption in bullish / base / bearish with stated rationale cracks the box open: you'll find that valuation results are 80% determined by your growth and WACC assumptions, and tweaking one flips the conclusion. Peer comparison uses median rather than mean, because means get distorted by outliers. The cross-check step is critical: if multiples say cheap but DCF says expensive, it means your growth assumption may diverge from what the market is currently pricing--and that's exactly what you need to think through.
Expert: Risk Decomposition & Investment Thesis
When you've read the three statements and built a valuation, and want a full risk assessment and investment-thesis construction. One prompt chains four-dimension risk decomposition + bull/bear logic + sensitivity analysis--better than "tell me if this stock is worth buying" because it forces the AI to inspect dimension by dimension and produce falsifiable logic rather than fuzzy conclusions.
You are a senior buy-side analyst. Based on the company, financials, and industry information I provide, output:
1. Four-dimension risk decomposition:
- Operating risk: moat of the core business, customer / supplier concentration, capacity utilization, product lifecycle
- Financial risk: leverage level, short-term solvency (current ratio / quick ratio), cash flow stability, goodwill impairment risk
- Industry risk: competitive landscape shifts, regulatory policy, technology disruption, cycle position
- Governance risk: ownership structure, related-party transactions, management integrity & incentives, disclosure quality
2. Bull / bear logic:
- Bull case (3 points): each with supporting evidence and a falsifiable test condition
- Bear case (3 points): each with supporting evidence and a falsifiable test condition
3. Sensitivity analysis:
- Identify the 3 variables with the greatest impact on valuation / conclusions
- For each, show the effect of ±10% / ±20% on key metrics
4. Investment thesis summary:
- One paragraph: what is the core tension, under what conditions does the bull case hold, under what conditions does the bear case hold
- List 3-5 key metrics / events to track going forward
Constraints:
1. Do not fabricate market data, stock-price movements, or regulatory document content; for anything not provided, mark "to be supplemented"
2. Bull/bear logic must include falsifiable test conditions; "promising industry outlook" type statements that cannot be verified are not allowed
3. Sensitivity analysis must be derived from the financial data already provided; do not speculate on missing data
4. All conclusions must note "relies on information provided; actual conditions require manual verification against authoritative sources"
5. Avoid empty evaluative language; every conclusion must be backed by a specific number or fact
6. End with: "The above is AI-assisted analysis, not investment advice, and must be reviewed by a qualified professional in compliance with applicable financial regulations"
Company + financials + industry: {{paste company overview, key financials, industry information}}Why four dimensions instead of "analyze the risk": the four-dimension structure forces the AI to cover each in turn, so it can't skip governance--the dimension most easily overlooked but often the most damaging. Many blowups showed clear governance risk well before the financials imploded (abnormal related-party transactions, declining disclosure quality). Requiring "falsifiable test conditions" on bull/bear logic prevents the AI from writing correct-but-useless boilerplate like "promising industry, strong competitiveness": every thesis must be falsifiable by a specific metric or event, e.g., "Bull case: new product ramp drives revenue growth back to 20%+; test condition: does new-product revenue share exceed 15% in the next two quarters." Sensitivity gets its own step to quantify "which assumption is most lethal"--if the valuation is extremely sensitive to WACC but not to growth, you know to spend your effort pinning down the discount rate.
Cheatsheet: Understand a Financial Report in 30 Seconds
When you don't have time for a full three-statement extraction and just want to quickly judge whether a report looks "healthy or suspicious." One compact prompt checks five of the most critical signals.
You are a senior financial analyst. Quickly scan this financial report and judge the company's financial health using 5 signals:
1. Earnings quality: is net income and operating cash flow moving in the same direction? (both up = healthy; profit up / cash flow down = warning)
2. Revenue quality: accounts receivable growth vs. revenue growth (receivables growing far faster = loosening credit to boost sales, warning)
3. Asset quality: inventory growth vs. revenue growth (inventory piling up = slow sales or impairment risk)
4. Debt pressure: short-term borrowings vs. cash & equivalents (short-term debt far exceeding cash = liquidity strain)
5. Non-recurring items: adjusted net income vs. net income (large gap = earnings rely on non-recurring items, unsustainable)
Output: for each signal, give a "healthy / neutral / warning" judgment + a one-line reason + the corresponding numbers.
Constraints: judge only based on the financial report provided; do not fabricate data; avoid empty evaluative language; end with "The above is AI-assisted analysis, not investment advice, and must be reviewed by a qualified professional."
Financial report text / data: {{paste}}This prompt's value is that its five signals map to the five most common financial-report window-dressing techniques: profit-cash flow divergence is the classic--profitable on paper but not collecting cash; receivables and inventory outpacing revenue signal credit loosening and channel stuffing; a high non-recurring items ratio means the core business isn't making money. Using "healthy / neutral / warning" tiers instead of a score is deliberate: a quick check doesn't need a precise rating, it just needs to tell you "should I dig deeper."
Universal Constraints & Compliance
Regardless of which level you use, the following constraints are embedded in every prompt:
- No fabricated data: The AI must not fabricate stock prices, market caps, latest financial data, or regulatory document content. All numbers must be manually verified against authoritative sources (exchange filings, audit reports, data terminals like Wind / Bloomberg). The AI may misread rows, mix up periods, or transpose figures.
- Separate AI analysis from human review: Every prompt ends with "The above is AI-assisted analysis, not investment advice, and must be reviewed by a qualified professional." AI output is structured organization and reasoning assistance, not investment decisions.
- Compliance baseline: Financial analysis involves investment decisions. AI output is not investment advice; rely on authoritative financial reports and data sources, and comply with applicable financial regulations. In many jurisdictions, providing investment advice requires a license--treat AI analysis as a research aid, not as investment advice to be published.
- De-AI-flavor tips: AI-written financial analysis tends to pile on filler like "in summary," "it is worth noting," and "the company has broad prospects." Adding a constraint to your prompt--"avoid empty evaluative language; every conclusion must be backed by a specific number or fact"--effectively suppresses the AI flavor.
Recommended models: financial reports often run over a hundred pages. For long-document parsing, Claude (200K context, stable on tables and footnotes) or DeepSeek (long reasoning chains, suited to valuation calculations and logical deduction) are recommended. Both perform reasonably on number extraction and logical reasoning, but critical numbers must still be manually verified.
Sources
- Anthropic Prompt Engineering overview (structured, step-by-step, context-anchored principles): https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview
- Damodaran valuation methodology (DCF, multiples, risk assessment--the standard framework, NYU Stern): https://pages.stern.nyu.edu/~adamodar/
- CFA Institute financial statement analysis framework (three-statement analysis, ratio analysis, risk identification): https://www.cfainstitute.org/
- IFRS Foundation (International Financial Reporting Standards, line-item definitions and disclosure requirements): https://www.ifrs.org/
- OpenAI Prompt Engineering guide (clear instructions, split into steps, provide examples): https://platform.openai.com/docs/guides/prompt-engineering