SEC Filing Evidence Prioritization Prompt Template

This prompt is designed for financial analysis and investor-relations AI tools that need to extract and rank evidence from SEC filings. The core job-to-be-done is prioritizing risk factors, Management Discussion & Analysis (MD&A) statements, and footnote disclosures by their quantitative materiality and temporal proximity. The ideal user is an engineering lead or AI builder integrating this into a product that serves analysts, portfolio managers, or compliance officers who require a structured, defensible evidence set, not just a text summary.

Use this prompt when your application must produce a disclosure-priority evidence set from one or more SEC filing documents (10-K, 10-Q, 8-K, etc.). It is appropriate when the downstream task is risk assessment, investment memo drafting, or financial due diligence. The prompt requires a specific [FILING_TEXT] input and works best when you also provide a [REPORTING_DATE] to anchor temporal relevance. Do not use this prompt for general financial news summarization, real-time market data interpretation, or as a replacement for a qualified financial analyst's judgment. It is not designed for non-SEC documents like earnings call transcripts or press releases without adaptation.

The primary constraint is distinguishing specific, company-written disclosure language from generic, boilerplate regulatory text. A successful implementation must include an evaluation step that penalizes the selection of boilerplate risk factors (e.g., 'economic conditions may adversely affect our business') unless they are accompanied by specific, quantified impact statements. The prompt is a component in a larger grounding pipeline; it assumes you have already retrieved or segmented the relevant filing sections. It does not perform document retrieval itself. Before integrating, ensure your system can handle the large context windows required for full filing sections and that you have a clear schema for the output evidence objects.

Next, copy the prompt template and adapt the placeholders to your specific filing types and output requirements. Pay close attention to the [MATERIALITY_THRESHOLD] and [TEMPORAL_WEIGHT] variables, as these directly control the ranking behavior and must be tuned for your use case.

The following template is designed to be dropped into a financial analysis or investor-relations AI workflow. It instructs the model to rank evidence passages extracted from SEC filings—such as risk factors, MD&A sections, and footnote disclosures—by their financial materiality and recency. The prompt uses square-bracket placeholders for all dynamic inputs, making it straightforward to integrate into a retrieval-augmented generation (RAG) pipeline or a document intelligence harness.

textCopy
You are a financial evidence analyst. Your task is to prioritize excerpts from SEC filings for a downstream analysis or answer-generation step.

You will receive a set of passages extracted from one or more SEC filings. Each passage includes a source label, a filing date, and the extracted text.

[INPUT]

For each passage, assign a priority score from 1 (lowest) to 5 (highest) based on the following criteria:
- Quantitative materiality: Does the passage contain specific dollar amounts, percentage changes, or other quantified financial impacts?
- Temporal proximity: How recent is the disclosure relative to the current analysis date? More recent disclosures generally carry higher priority.
- Disclosure specificity: Does the passage contain company-specific, non-boilerplate language, or is it generic risk-factor language?
- Regulatory prominence: Is the passage from a section with higher disclosure prominence, such as Item 1A Risk Factors, Item 7 MD&A, or a footnote to the financial statements?

[CONSTRAINTS]
- Do not invent or assume financial figures not present in the provided text.
- If a passage is entirely boilerplate with no company-specific quantification, assign it a priority of 1.
- If two passages contain conflicting information, flag the conflict in your reasoning but do not resolve it.
- If a passage references a date more than two fiscal years before the most recent filing date in the input, reduce its temporal proximity score by at least 1 point.

[OUTPUT_SCHEMA]
Return a JSON array of objects with the following fields:
- passage_id: string (the identifier from the input)
- priority_score: integer (1-5)
- materiality_rationale: string (one sentence explaining the score, citing specific figures or language from the passage)
- boilerplate_flag: boolean (true if the passage is primarily boilerplate disclosure)
- conflict_flag: boolean (true if this passage conflicts with another passage in the set; include the conflicting passage_id in conflict_with)
- conflict_with: string or null (passage_id of the conflicting passage, if any)

[EXAMPLES]
Example input:
[
  {
    "passage_id": "p1",
    "source": "10-K FY2024 Item 7",
    "filing_date": "2025-02-15",
    "text": "Revenue increased 12% to $4.2 billion, driven primarily by higher subscription volumes in North America."
  },
  {
    "passage_id": "p2",
    "source": "10-K FY2024 Item 1A",
    "filing_date": "2025-02-15",
    "text": "Our business is subject to risks associated with economic downturns, which could materially and adversely affect our results of operations."
  }
]

Example output:
[
  {
    "passage_id": "p1",
    "priority_score": 5,
    "materiality_rationale": "Contains specific 12% revenue growth and $4.2B figure with geographic driver detail.",
    "boilerplate_flag": false,
    "conflict_flag": false,
    "conflict_with": null
  },
  {
    "passage_id": "p2",
    "priority_score": 1,
    "materiality_rationale": "Generic risk-factor language with no company-specific quantification.",
    "boilerplate_flag": true,
    "conflict_flag": false,
    "conflict_with": null
  }
]

To adapt this template for your own pipeline, replace the [INPUT] placeholder with your retrieved passage set, including passage IDs, source labels, filing dates, and extracted text. Adjust the [CONSTRAINTS] block to reflect your specific materiality thresholds, temporal windows, or regulatory filing types. The [OUTPUT_SCHEMA] can be extended with additional fields such as section_type, filing_type, or confidence_score if your downstream application requires them. Always validate the output against the schema before passing it to an answer-generation or summarization step.

Before deploying, run this prompt against a golden dataset of SEC filings that includes both boilerplate and specific disclosure language. Measure whether the model correctly assigns low priority to generic risk factors and high priority to quantified MD&A statements. If the model over-prioritizes recent but immaterial disclosures, tighten the quantitative materiality criterion in the [CONSTRAINTS] block. For high-stakes financial analysis workflows, route priority-5 passages through a human review step before they influence investment conclusions or public-facing outputs.

This prompt is designed to operate as a post-retrieval ranking and filtering step inside a larger RAG pipeline, not as a standalone chatbot. The typical flow is: (1) user query or analysis task arrives, (2) a retrieval system pulls candidate passages from a SEC filing corpus (10-K, 10-Q, 8-K, S-1, proxy statements), (3) this prompt ranks and prioritizes those passages by quantitative materiality and temporal proximity, and (4) the ranked evidence set is passed to a downstream answer-generation or report-drafting prompt. The harness must enforce that the model never sees raw filings directly; it only sees pre-retrieved passages plus filing metadata.

Input assembly requires three structured fields injected into the prompt template: [FILING_METADATA] (ticker, filing type, period end date, filing date), [RETRIEVED_PASSAGES] (an array of objects with passage_id, section_type, text, and page_number), and [ANALYSIS_OBJECTIVE] (the user's task, e.g., 'Identify top 5 risk factors for revenue concentration'). The harness should validate that section_type is one of the SEC-standard categories—Risk Factors, MD&A, Footnotes, Legal Proceedings, Quantitative Disclosures—before the prompt is assembled. Reject passages with missing or unrecognized section types rather than letting the model guess. The [OUTPUT_SCHEMA] placeholder should be replaced with a strict JSON schema requiring ranked_evidence (ordered array of passage IDs with priority scores 0-1 and materiality flags), boilerplate_flags (boolean per passage indicating generic vs. specific disclosure language), and temporal_proximity_days (days between passage reference date and filing date).

Model choice and tool use: Use a model with strong structured output support (GPT-4o, Claude 3.5 Sonnet, or equivalent) and enable JSON mode or structured outputs with the schema above. This prompt does not require tool calls, but the harness should wrap it in a retry loop with schema validation. If the output fails JSON schema validation, retry once with the validation error appended to [CONSTRAINTS]. If the second attempt fails, log the failure and escalate to a human review queue rather than silently falling back to unstructured text. For high-stakes financial analysis, never skip validation—a malformed evidence ranking can propagate materiality errors into downstream reports.

Logging and observability: Log every invocation with prompt_version, model_id, filing_accession_number, passage_count, output_ranked_count, boilerplate_ratio (fraction of passages flagged as boilerplate), and latency_ms. This data is essential for detecting prompt drift when the SEC updates filing requirements or when a new model version changes ranking behavior. Set an alert if boilerplate_ratio exceeds 0.7, as this indicates the retrieval step is pulling too many generic risk-factor paragraphs and needs tuning. For audit trails, store the full ranked evidence set alongside the downstream generated output so that compliance reviewers can trace every claim back to its source passage.

Human review gates: If the prompt's [RISK_LEVEL] is set to high (e.g., material weakness analysis, going-concern assessment, or fraud risk factor review), the harness must route the ranked evidence set to a human analyst for approval before it feeds into any answer-generation step. The review interface should display each ranked passage with its section type, materiality flag, and boilerplate status, and allow the analyst to re-rank, remove, or annotate passages. Only after approval should the evidence set proceed. For medium risk, log and sample 10% of outputs for retrospective human review. For low risk (e.g., general industry comparison), automated flow is acceptable but still requires the observability logging described above.

What to avoid: Do not use this prompt as a single-step answer generator—it ranks evidence, it does not synthesize answers. Do not skip the boilerplate detection eval; SEC filings are dense with legally required generic language that looks relevant but carries no company-specific signal. Do not treat all sections as equally authoritative; the harness should weight Risk Factors and MD&A higher than Exhibit indices or signature pages by pre-filtering retrieval candidates before they reach this prompt. Finally, do not deploy without a temporal freshness check: if the filing date in [FILING_METADATA] is more than 90 days old for a quarterly analysis task, the harness should warn the user and suggest retrieving more recent filings before proceeding.