An AI Prompt to Analyze a 10-K Into a Forensic Checklist

Reading a 10-K cover to cover is a job. The risk factors run twenty pages, half of it boilerplate that's been copied forward for years. The MD&A buries the one sentence that matters under careful hedging. So people paste the section into ChatGPT and ask it to "summarize the risks." It returns a tidy paragraph, and that paragraph is exactly the problem: it flattens the new risk and the recycled one into the same calm prose. A forensic 10-K analysis prompt does the opposite. It extracts, ranks, and flags, and it ties every finding to a quote you can check.

The shift is from summary to structure. Instead of "tell me about the risks," the prompt asks: which risk factors are new this year, where did the accounting language change, where does management's tone diverge from the numbers, and what's the source quote for each. That output is a checklist, not an essay. And it's only useful if every line cites the filing, because the model will otherwise paraphrase a figure wrong and you'd never know.

Why "summarize this 10-K" guides leave the rigor out

The pages that rank for 10-K analysis are mostly generic summarize-and-go walkthroughs, often written for an audience that isn't doing financial diligence at all. Seer Interactive's guide (seerinteractive.com) is aimed at marketers tagging opportunities with hashtags like #SEO; it treats the 10-K as a strategy source and includes no accounting red flags, no verification step, and no responsible-use note. The Endgame walkthrough (endgame.io) is built for sales reps researching prospects: upload the filing, ask broad questions across seven topics, done. No forensic framework, and only a passing mention of asking for citations.

Both are fine for what they are. Neither gives you a forensic output contract. There's no structured extraction of new-versus-recycled risk, no place where the prompt cross-checks management's narrative against the financial statements, and no honest disclaimer that the output needs verification and isn't investment advice.

That's the winnable gap, and it matters because the stakes are real. A hallucinated figure in a marketing summary is harmless. A hallucinated figure in a risk assessment is dangerous.

What you can do with this prompt

• Extract risk factors and rank them by materiality instead of reading order
• Flag which risks are new this year versus copied forward as boilerplate
• Surface accounting red flags: changed policies, revenue-recognition shifts, unusual reserves
• Pull MD&A signals where management tone diverges from the reported numbers
• Tie every single finding to a quoted passage you can verify on EDGAR
• Compare two filings on an identical checklist instead of two different summaries

Anatomy of the 10-K analysis prompt

The forensic version has a pasted section, a fixed checklist contract, and a hard citation rule.

Variables
  {{filing_section}}   → pasted text: Risk Factors, or MD&A, or both
  {{prior_year_section}} → optional: same section from last year's 10-K
  {{company_name}}     → for labeling only, never as a fact source

Prompt
  Role: forensic financial-disclosure analyst
  Task: produce the fixed checklist below from {{filing_section}} ONLY
  Rule: every finding MUST quote the source sentence verbatim
  Rule: if a section is not present in the pasted text, write
        "Not in provided text" — never analyze from outside knowledge

Output contract (fixed checklist)
  1. Top 5 material risks — ranked, each with a verbatim quote
  2. New vs. recycled — risks new this year vs. carried forward
       (only if {{prior_year_section}} is provided)
  3. Accounting red flags — policy changes, reserve/recognition shifts
  4. MD&A signals — where narrative tone diverges from the figures
  5. Verification queue — every figure to check against the statements

The "from the pasted text ONLY" rule is what keeps it forensic. Models carry a lot of general knowledge about big public companies, and without that boundary they'll blend the actual filing with half-remembered facts from training. You want analysis of this document, not a vibe about the company.

The stance: citations or it didn't happen

The single rule that separates a useful 10-K prompt from a liability is the verbatim-quote requirement. Make every finding carry the source sentence. It does two things. It forces the model to ground each claim in the text instead of generating a plausible risk, and it gives you a fast verification path: skim the quotes against the filing, and anything the model invented has no quote to stand on. A finding without a quote is a finding to delete. That's not optional rigor for financial work. It's the floor.

Step-by-step usage

1. Gather inputs

Pull the filing from SEC EDGAR, not a summary site. Copy the Risk Factors section, the MD&A, or both into plain text. If you want the new-versus-recycled comparison, grab last year's matching section too.

2. Fill variables

Paste the section into {{filing_section}}. Drop the prior year into {{prior_year_section}} if you have it. The {{company_name}} is a label only; the prompt is told never to treat it as a fact source, which stops the model from leaning on what it "knows" about the company.

3. Run the prompt

Paste into Claude or ChatGPT. Claude handles long pasted filings better in a single context and tends to keep quotes verbatim; GPT-4o is reliable on shorter sections but will sometimes lightly paraphrase a "quote," so the verification step isn't skippable on either.

4. Post-process

Work the Verification queue line by line against the actual financial statements. Every figure the model surfaced, every quote, checked against EDGAR. This is the non-negotiable part. The prompt is a reading aid that points you at what to scrutinize; it does not replace reading the numbers.

5. Iterate

Run the same prompt on a peer company's filing and you get the identical checklist, which is what makes comparison possible. Two summaries are apples and oranges. Two checklists with the same five sections line up.

Prompt-craft patterns that hold

Closed-world instruction. "Analyze only the text provided; if something isn't in it, say so" turns off the model's general knowledge. For a 10-K that's essential, because the blend of training data and pasted text is invisible in the output and impossible to audit after the fact.

Use ONLY the text in {{filing_section}}.
Do not use any knowledge about {{company_name}} from outside this text.
If a checklist item has no support in the text, write "Not in provided text".

Quote-then-claim ordering. Ask for the verbatim quote first, then the analysis. "Quote the sentence, then state the red flag it implies" produces grounded findings. The reverse order lets the model write the conclusion first and reverse-engineer a quote, which is how paraphrase creep starts.

The investment-advice boundary. End the prompt with a refusal line: "Do not give buy/sell/hold recommendations; this is disclosure analysis only." The model will otherwise drift toward a verdict, and a 10-K prompt that outputs a stock opinion is exactly what you don't want feeding a real decision.

Variables you'll set

Getting started

1. Pull the filing from SEC EDGAR and copy the section as plain text.
2. Paste it in with the closed-world rule and the verbatim-quote requirement.
3. Add last year's section if you want new-versus-recycled flagging.
4. Run it in Claude or ChatGPT, keeping the section short enough to stay in context.
5. Work the Verification queue against the actual financial statements.
6. Delete any finding that doesn't carry a source quote.
7. For the packaged version with the forensic checklist and citation rule already built in, use the 10-K Filing Forensics Prompt Pack.

Forensic reading is a research discipline as much as a finance one. If your analysis rolls up into a board update, the Board Pack Production Prompt Pack takes the same source-or-silence rigor into the deck narrative.

A 10-K prompt is only ever a reading aid; the filing is the source of truth and the output is analysis support, not investment advice. Keep the human in the loop on every figure. For the discipline of catching invented facts in any AI output, see the hallucination faithfulness check, and for rolling forensic findings into a leadership narrative, the board deck narrative prompt shows the same cite-or-omit pattern at deck level.