Watermark Identification Prompt

This prompt is designed for document processing teams who need to programmatically detect and classify watermarks that interfere with downstream text extraction, optical character recognition (OCR), or document classification tasks. The ideal user is an engineer or technical operator building a document ingestion pipeline who must distinguish between a genuine watermark and other artifacts like bleed-through from the reverse page, scanner noise, or page creases. The job-to-be-done is not just watermark detection, but producing a structured, machine-readable assessment of the watermark's type, opacity, location coverage, and a severity rating that quantifies how much it degrades text extraction quality.

Use this prompt when you have a pre-processed image of a single document page and you need a structured decision about whether a watermark is present and how it affects the page. This is a classification and assessment task, not a removal task. The prompt requires a clear image input and works best when the document has already been separated into individual pages. It is not suitable for real-time video streams, multi-page document-level summaries without per-page analysis, or for directly removing the watermark from the image. For high-stakes compliance or legal workflows, the output must be treated as a signal for a human review queue, not as a final automated decision. The severity rating should be used to route pages: low-severity pages can proceed to extraction, while high-severity pages are flagged for manual review or specialized image preprocessing.

Before integrating this prompt, ensure your pipeline can isolate single-page images and that you have defined clear thresholds for the 'interference severity rating' that align with your downstream task's tolerance for error. A common failure mode is deploying this prompt on documents with heavy bleed-through from the reverse side, which can be misclassified as a faint watermark. To mitigate this, pair this prompt with a pre-processing step that identifies and labels potential bleed-through artifacts, and use the eval checks provided in the full playbook to measure false-positive rates on a golden dataset of known artifacts. The next step is to copy the prompt template and wire it into a validation harness that checks the output schema before the result is used for routing.

The following prompt template is designed to be copied directly into your application code, test suite, or prompt management system. It uses square-bracket placeholders for all dynamic inputs—such as the document image, output schema, and risk tolerance—so you can wire it into a document processing pipeline without rewriting the core instruction. The template assumes the model receives a page image and returns structured JSON describing each detected watermark's type, opacity, location coverage, and interference severity.

textCopy
You are a document forensics assistant specialized in watermark detection. Your task is to analyze the provided page image and identify all watermarks present. A watermark is any intentional, semi-transparent marking overlaid on or embedded in the document content, including text watermarks (e.g., "DRAFT", "CONFIDENTIAL"), image watermarks (e.g., logos, seals), and repeating pattern watermarks (e.g., diagonal "VOID" patterns).

[INPUT]
- Page image: [IMAGE]
- Page number in document: [PAGE_NUMBER]
- Document type: [DOCUMENT_TYPE]

[CONTEXT]
- Known watermarks for this document source: [KNOWN_WATERMARKS]
- Previous page watermark state: [PREVIOUS_PAGE_STATE]

[OUTPUT_SCHEMA]
Return a JSON object with the following structure:
{
  "page_number": integer,
  "watermarks_detected": boolean,
  "watermarks": [
    {
      "watermark_id": "string (unique identifier for this watermark instance)",
      "type": "text | image | pattern",
      "content": "string (the text content for text watermarks, description for image/pattern)",
      "opacity_assessment": "faint | moderate | bold",
      "location_coverage": {
        "region": "full_page | header | footer | center | diagonal | scattered",
        "estimated_coverage_percent": number (0-100),
        "bounding_box_approximation": "string (e.g., 'top-left quadrant', 'full page overlay')"
      },
      "interference_severity": "none | low | moderate | high | blocking",
      "interference_details": "string (explain what text or content is obscured)",
      "confidence": number (0.0-1.0),
      "distinguishing_from_artifacts": "string (explain why this is a watermark and not bleed-through, scanner noise, or a pre-printed form element)"
    }
  ],
  "false_positive_risks": [
    {
      "region": "string",
      "appearance": "string",
      "why_excluded": "string (e.g., 'bleed-through from reverse page', 'scanner artifact', 'staple mark')"
    }
  ],
  "overall_assessment": "string (summary of watermark impact on text extraction for this page)"
}

[CONSTRAINTS]
1. Distinguish watermarks from bleed-through by checking if the marking aligns with content from the reverse side of the page. Bleed-through typically mirrors reversed text or images from the other side.
2. Distinguish watermarks from scanner artifacts by checking for repeating patterns at fixed intervals, color channel noise, or edge distortions that are mechanical rather than document-intrinsic.
3. Do not classify pre-printed form elements (borders, field labels, instructional text) as watermarks.
4. If a watermark is faint and does not obscure underlying text, set interference_severity to "none" or "low". If text is unreadable beneath the watermark, set to "high" or "blocking".
5. For pattern watermarks, describe the pattern geometry (e.g., "diagonal lines at 45 degrees with 'VOID' text repeated every 2 inches").
6. If no watermarks are detected, return watermarks_detected: false with an empty watermarks array and explain in overall_assessment.

[RISK_LEVEL]
Current risk level: [RISK_LEVEL]
- If "low": proceed with extraction, flag watermarked regions for review only if interference_severity is "high" or "blocking".
- If "medium": route any page with watermarks_detected: true to human review queue.
- If "high": route all pages to human review regardless of watermark detection; this prompt output serves as an advisory note only.

[EXAMPLES]
Example input: Page with faint diagonal "DRAFT" pattern across center.
Example output:
{
  "page_number": 3,
  "watermarks_detected": true,
  "watermarks": [
    {
      "watermark_id": "wm_003_01",
      "type": "pattern",
      "content": "Diagonal 'DRAFT' text repeated at 45-degree angle, 1.5-inch spacing",
      "opacity_assessment": "faint",
      "location_coverage": {
        "region": "diagonal",
        "estimated_coverage_percent": 40,
        "bounding_box_approximation": "full page diagonal overlay"
      },
      "interference_severity": "low",
      "interference_details": "Underlying text remains legible; watermark is light gray and does not overlap dense text regions significantly",
      "confidence": 0.92,
      "distinguishing_from_artifacts": "Pattern is uniform and intentional; no alignment with reverse-page content; distinct from scanner noise which would show fixed-interval banding"
    }
  ],
  "false_positive_risks": [],
  "overall_assessment": "Single faint pattern watermark detected. Text extraction should proceed with minor accuracy impact in overlapping regions."
}

To adapt this template for your pipeline, replace each square-bracket placeholder with values from your document processing context. The [IMAGE] placeholder should receive a pre-processed page image (consider deskewing and contrast normalization before sending). The [KNOWN_WATERMARKS] field lets you pass prior knowledge about standard watermarks for this document source, reducing false positives. The [RISK_LEVEL] parameter controls downstream routing behavior directly in the prompt, so your application can read the output and act without additional classification logic. Before deploying, run this prompt against a golden dataset containing known bleed-through cases, scanner artifacts, and genuine watermarks to calibrate your confidence thresholds and verify that false_positive_risks entries are populated correctly.

The Watermark Identification Prompt is designed to be called as a dedicated analysis step within a document ingestion pipeline, not as a standalone chat interaction. It should receive a pre-processed page image or region crop along with metadata about the document's origin (scanned, born-digital, photographed). The prompt's primary job is to produce a structured classification of any watermark present, including its type, opacity, location coverage, and an interference severity rating that downstream extraction steps can use to adjust their confidence or abort processing on severely obstructed pages.

Wire this prompt into your application by first isolating candidate pages using a lightweight pre-filter. A computer vision heuristic—such as detecting low-contrast repeating patterns or large semi-transparent text regions—can reduce the number of pages sent to the model, controlling latency and cost. For each candidate page, call the prompt with the page image and a strict output schema. Parse the JSON response and validate that watermark_type is one of the expected enum values (text, image, pattern, none). If interference_severity is high or critical, route the page to a specialized OCR pipeline or flag it for human review rather than feeding it into your standard text extraction flow. Log the location_coverage bounding boxes so downstream systems can mask or ignore those regions.

Failure modes in production cluster around false positives from scanner artifacts and bleed-through. Implement a validation layer that cross-references the model's output with page-level metadata. If the model reports a watermark on a born-digital PDF page, flag the result for review—true watermarks are rare in native digital documents. For scanned documents, compare the detected watermark location against the reverse side of the page when duplex scanning is available; overlapping text regions suggest bleed-through rather than an actual watermark. Build a retry loop with a temperature of zero for consistency, but limit retries to two attempts—if the model returns inconsistent watermark_type values across retries, log the ambiguity and default to a conservative interference_severity of medium to avoid silent data loss.

Model choice matters here. Vision-capable models with strong document understanding (such as GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro) handle this task well. Avoid text-only models that require pre-extracted OCR text as input—watermarks by definition interfere with OCR quality, so relying on OCR output creates a circular dependency. For high-throughput pipelines processing thousands of pages, consider batching multiple page crops into a single request where the model can assess each independently, but keep batch sizes small (3–5 pages) to maintain output reliability. Always include a human review queue for pages where interference_severity is critical and the document is part of a compliance or legal workflow.