On July 30, 2025, Microsoft filed its fiscal 2025 10-K. Inside was 281.7 billion in annual revenue, a 15% year-over-year jump, and a capex line that had quietly grown to 88.7 billion. Most retail investors never opened the document. A handful who did, armed with Claude AI and a structured prompt, pulled out the three things that actually mattered in under ten minutes.
A 10-K is the most information-dense document a public company produces. It is also 150 to 300 pages long, written by lawyers, and optimised to be legally defensible rather than readable. That combination is exactly the problem Claude AI is good at solving. Give it a clear structure, and it becomes a filter that catches everything a press release is designed to hide.
This guide gives you a 6-section prompt framework that turns any 10-K into a one-screen investable read. You will see it run live on Microsoft's FY2025 filing, then applied to Meta's FY2024 filing so you can compare outputs side by side. The framework works on any US-listed company, and transfers to non-US annual reports with small tweaks.
What Is a 10-K, and Why Use Claude AI to Read It?
A 10-K is the annual report every US-listed public company files with the SEC. It is the comprehensive, audited, legally binding document that covers the business description, risk factors, financial statements, management's discussion and analysis (MD&A), and disclosures a press release leaves out. European and UK equivalents are the Annual Report and Accounts. Most other markets have their own annual report format that follows roughly the same structure.
Using Claude AI to read a 10-K means asking it to do three things humans do badly at scale: find the five sentences that actually matter in MD&A, translate lawyered risk language into plain English, and compare this year's numbers to last year's across fifteen segment lines. Claude does not replace the filing. It gives you a reliable index into it, so you know exactly which 5 pages of the 300 are worth your own time.
Think of the 10-K like a detailed property survey on a house you are about to buy. You could read every page yourself. But most people hire an inspector who knows where the real problems hide. Claude is the inspector. The survey still matters, you still skim it at the end, and you still make the decision.
Why This Matters More Than a Press Release
Press releases are designed to make the quarter look good. 10-Ks are designed to satisfy the SEC. The gap between the two is where every major short thesis, accounting scandal, and hidden compounding advantage has ever been documented in plain sight. Enron, Wirecard, and Valeant were all visible in the 10-K years before they blew up. Apple's services margin was visible in the 10-K for years before the market priced it in.
$88.7 billion in FY25 capex, a 59% year-over-year jump, almost entirely AI datacenters, GPUs, and networking.
Three things make 10-K reading higher-value today than even five years ago. Segment disclosure rules have tightened, so you get more granular margin data per business line. AI and infrastructure capex has become the single biggest driver of mega-cap valuation and sits primarily in the 10-K notes, not the press release. And filings have become longer, not shorter, which widens the reading gap between people who have a framework and people who do not.
Figure 1: The 6-section framework applied to Microsoft's FY2025 10-K, live in Claude with web search on.
How to Read a 10-K with Claude AI: The 6-Section Framework
The prompt below is the exact one used to produce every example that follows. Copy it into Claude, swap the ticker and fiscal year, and you have a repeatable 10-K read that takes under ten minutes.
Step 1: Frame Claude as an Equity Analyst
Open with "You are an equity analyst." This is not cosmetic. The analyst framing forces Claude to prioritise materiality, variance versus the prior year, and MD&A tone over generic summary prose. Without it you get a press-release rewrite.
Step 2: Ask for a 6-Section Scannable Read
Request exactly these six sections in order: (1) a three-sentence business snapshot including segment revenue mix, (2) the top three year-over-year quantitative changes with exact figures, (3) the top three risk factors translated out of legal-speak, (4) what MD&A reveals that the press release does not, (5) one red flag most investors will miss, (6) a one-line thesis impact statement. The order and count are deliberate. They force Claude to produce comparable output for every company, which is what makes the framework useful across time and across stocks.
Step 3: Enable Web Search for the Most Recent Filing
For filings published in the last 12 months, turn on Claude's web search. Without it, Claude may rely on training data and miss the most recent annual filing. With it, Claude pulls directly from SEC EDGAR, the company investor relations page, and the press coverage around the filing, and cites sources inline. For filings more than a year old, training data alone is usually fine.
Step 4: Verify the Top Three Numbers Against the Source
Claude will return three exact figures in section 2. Open the 10-K PDF or the company's 8-K press release and check those three numbers. If any one is wrong, everything downstream is suspect. This is the step retail investors skip most often and the one that turns the framework from an AI demo into an actual research process.
Step 5: Re-Run the Same Prompt on a Comparable Company
The real value is comparability. Run the identical prompt on one competitor from the same cycle. Below you will see the six-section output for Microsoft's FY2025 10-K followed by the same six sections for Meta's FY2024 10-K. Stacking the two instantly shows where the AI capex arms race has shifted and where margin trajectory is diverging.
Real Examples: MSFT FY2025 vs META FY2024
The outputs below are directly from Claude. They have been lightly cleaned for readability but not rewritten. The same six-section prompt produced both.
Example 1: Microsoft FY2025 10-K, filed July 30, 2025
Business snapshot: Microsoft sells cloud infrastructure (Azure), commercial productivity software (Microsoft 365, Copilot, Dynamics, LinkedIn), and personal computing (Windows OEM, Surface, Xbox, Bing). FY25 total revenue of $281.7 billion splits across three segments: Productivity and Business Processes ~43% ($120.8B), Intelligent Cloud ~38% ($106.3B), and More Personal Computing ~19% ($54.6B). The real growth engine is Azure, which crossed $75 billion for the first time, up 34%.
Year-over-year changes: Revenue $281.7B versus $245.1B prior year (+15%). Operating income $128.5B, up 17%, with operating margin expanding ~100 bps to about 45.6% despite massive reinvestment. Capex (PP&E plus finance leases) hit $88.7 billion for FY25 versus roughly $55.7B in FY24, a ~59% jump. Commercial remaining performance obligation (contracted but not yet delivered revenue) reached $368 billion, giving multi-year visibility into the cloud book.
Risks in plain English: infrastructure overbuild risk (short-lived GPU assets at $88B per year), OpenAI concentration risk (a single partner underpins a disproportionate share of the AI narrative), and AI liability risk (Copilot-era monetisation is moving faster than the regulatory and liability framework).
What MD&A reveals: an August 2024 segment recast moved commercial Microsoft 365 into Productivity and Business Processes. Headline segment growth therefore looks cleaner because of the reshuffle, not just organic strength. FASB updates added segment-level cost and operating expense disclosures, making per-segment margins visible in a way they were not before. Azure was disclosed in dollars, not just growth percent, reversing years of opacity.
Red flag most investors will miss: the tight capex-backlog coupling. Management explicitly ties $88B capex to the $368B backlog. If bookings decelerate, capex has to follow, and Microsoft becomes the swing factor for global AI infrastructure demand. One-line thesis impact: long-term compounder thesis intact, but at current capex intensity and OpenAI concentration this is a high-conviction AI infrastructure bet, not a low-risk software annuity.
Example 2: Meta Platforms FY2024 10-K, filed January 29, 2025
Same prompt. Completely different shape of business.
Business snapshot: Meta sells digital advertising on Facebook, Instagram, Messenger, WhatsApp, and Threads, plus consumer hardware (Quest, Ray-Ban Meta glasses) and metaverse investment via Reality Labs. FY24 total revenue of $164.5 billion splits into just two reported segments: Family of Apps ~98.7% ($162.4B, almost entirely advertising) and Reality Labs ~1.3% ($2.1B). Geographic mix is re-accelerating abroad: US and Canada grew 18% in 2024, Europe 26%, Asia-Pacific 22%, and Rest of World 31%.
Year-over-year changes: Revenue $164.5B versus $134.9B (+22%). Operating income $69.38B, up 48%, lifting operating margin from ~35% to ~42%, a ~750 bps expansion driven by 10% average price per ad growth plus 11% impression growth. Capex $39.23 billion for FY24, up from ~$28B in FY23 (~40% jump), and guided to $60-65 billion for FY25. Reality Labs operating loss deepened to $17.73B from $16.1B in FY23, with management explicitly guiding further increases in 2025.
Risks in plain English: regulatory and ad-tech signal loss (privacy rules in the EU, UK, and multiple US states continue to compress targeting precision), Reality Labs cash burn risk (cumulative losses since 2020 now exceed $65B with no production-scale commercial hardware), and AI infrastructure commitment risk (capex stepping to $60-65B in FY25 is a bet that Llama-driven ad monetisation arrives in time).
What MD&A reveals: total expenses guidance for 2025 is $114-$119 billion, and infrastructure-related costs are the single largest driver. Meta explicitly disclosed that impression growth is being sustained by Reels monetisation improvements, not just overall user growth, which is a quieter signal than the revenue headline suggests.
Red flag most investors will miss: the capex-to-revenue ratio. Meta capex of $39B on $164.5B revenue is ~24%, and the FY25 guide pushes that above 35% at the midpoint. That is closer to a semiconductor fab than a consumer internet business. One-line thesis impact: the ad engine is performing better than bulls expected, but at this capex trajectory Meta is effectively two companies, a high-margin ad business funding an escalating AI and Reality Labs cash burn.
Two 10-Ks, one prompt, comparable six-section outputs in under twenty minutes total.
Side-by-Side: MSFT FY25 vs META FY24
Stacking the outputs is where the comparability actually pays off. A single table shows the two biggest AI-spending businesses on the planet with very different operating shapes.
- Metric: MSFT FY2025 | META FY2024
- Revenue: $281.7B (+15% YoY) | $164.5B (+22% YoY)
- Operating margin: ~45.6% (+100 bps) | ~42% (+750 bps)
- Capex: $88.7B (+59% YoY) | $39.2B (+40% YoY)
- Capex intensity (capex/revenue): ~31% | ~24% (FY25 guide: ~37%)
- Segment count: 3 | 2 (ad vs Reality Labs)
- Key backlog disclosure: $368B commercial RPO | Expenses guide $114-119B FY25
- Dominant risk: OpenAI + infra overbuild | Ad signal loss + RL burn
- Thesis shape: AI infrastructure bet | Ad engine + RL option
Figure 2: Stacked 6-section outputs for MSFT FY2025 and META FY2024, rendered in Claude.
Common Mistakes When Using Claude AI to Read 10-Ks
Mistake 1: Asking for a Summary Instead of a Structured Read
"Summarise Microsoft's 10-K" gets you a press release in paragraph form. The six-section structure is what forces Claude to pull risk factors, MD&A, and red flags, not just re-write the CEO letter. Always use the numbered framework.
Mistake 2: Trusting Exact Numbers Without Verification
Models occasionally produce figures that are close to correct but slightly off, especially on filings published within the last 30 days where search coverage is thin. Always verify the three headline numbers (revenue, operating income, capex) against the actual 8-K or 10-K PDF. Two minutes with the press release will save you a wrong thesis.
Mistake 3: Skipping the Prior-Year Delta
A 10-K read with no comparison to the prior year is a photograph. A 10-K read with year-over-year deltas is a trajectory. The prompt asks for three quantitative changes for a reason. If Claude returns point-in-time numbers only, re-prompt with "all numbers must be given year-over-year."
Mistake 4: Ignoring the Notes to the Financial Statements
Most of the value in a 10-K sits in Item 7 (MD&A) and Item 8 (notes to the financial statements), not the headline P&L. The framework explicitly asks what MD&A reveals beyond the press release, because that is where share buyback intent, capex allocation by segment, and tax-rate normalisation changes live.
Mistake 5: Running It Once a Year Only
10-K reading is maximum value when it is repeatable. Save every output. Re-run the same prompt a year later on the new filing. The delta between your FY24 output and your FY25 output on the same company is more valuable than either read in isolation. This is also true across competitors: the value of the comparison above is that both outputs share exactly the same six-section skeleton.
Frequently Asked Questions
How long does it take to run this framework on one 10-K?
With web search enabled, roughly 30-90 seconds for Claude to produce the six sections, plus 3-5 minutes of your own time to verify the three headline numbers against the source. Total: under 10 minutes per stock, which is about 30 times faster than reading cover to cover.
Does this work for non-US companies?
Yes, with one small change. Replace "10-K" with "annual report and accounts" or the local equivalent (20-F for non-US companies listed on US exchanges, for example). The six-section structure is format-agnostic. It relies on concepts (segment mix, YoY deltas, risk factors, MD&A-equivalent commentary) that every major market's annual report contains.
Can I use this for small-cap stocks?
Yes, and this is where the framework has the highest marginal edge. Big-cap 10-Ks are dissected by hundreds of sell-side analysts within 24 hours of filing. A small-cap industrial or a mid-cap regional bank may be covered by only two or three analysts, and a disciplined six-section read is genuinely incremental information.
Should I use Claude for the financial statements themselves?
Use Claude as an index, not a calculator. Ask Claude which line items changed materially year over year and where they live in the document. Then open the actual statements to verify. AI models are excellent at direction and prioritisation on financial data and merely acceptable at extracting specific cells. Treat them as a junior analyst who flags what to look at, not a senior analyst who produces the numbers.
What to Watch Next
- v Does Microsoft's FY26 capex exceed the $88.7B FY25 base, or does the commercial RPO growth rate decelerate first?
- v Does Meta's FY25 capex land inside the $60-65B guide, or does the upper bound move higher on the next earnings call?
- v Does Azure stay disclosed in dollars in future filings, or does Microsoft revert to percentage-only growth language?
- v Does Reality Labs loss peak in 2025 as guided, or widen again in 2026?
- v Do segment-level cost disclosures (new under FASB) expose margin gaps competitors start exploiting in FY26 filings?
Key Takeaways
- Use a fixed six-section prompt on every 10-K so outputs are directly comparable across companies and across years.
- Always start with the analyst persona. Generic summary prompts produce generic output that just rewrites the press release.
- Enable web search for filings less than 12 months old. Rely on training data only for older filings.
- Verify the top three headline numbers manually against the 8-K or 10-K PDF. Never skip this step.
- Force year-over-year deltas, not point-in-time numbers. Trajectory beats snapshot for investment decisions.
- Run the identical prompt on a competitor in the same cycle. The stacked view is where the actual edge sits.
- Save every output. The delta between this year and next year, on the same company, is the highest-value signal.