Multimodal Unsafe Content Detection Prompt

This prompt is designed for safety engineers and platform operators who need a single, unified classification layer for unsafe content across image, audio, and video inputs. Instead of running separate text-only classifiers on extracted transcripts or captions, this prompt instructs the model to reason directly over multimodal signals, produce modality-specific harm categories, and flag cross-modal correlations that single-modality systems miss. The primary job-to-be-done is to create a centralized, audit-ready safety screening decision that reduces the integration complexity and blind spots of stitching together multiple unimodal classifiers.

Use this prompt when your product accepts multimodal user uploads, when you are building a centralized safety screening API, or when you need audit-ready classification decisions with evidence pointers. It is appropriate for asynchronous or near-real-time screening workflows where a few seconds of latency is acceptable. The prompt requires that you provide the raw multimodal content (e.g., image bytes, audio waveforms, video frames) alongside any available metadata, and that you have defined a clear safety taxonomy with modality-specific harm categories. You must also configure a risk threshold for gating decisions (block, flag, allow, escalate) and have a human review queue for edge cases.

Do not use this prompt for real-time streaming video with sub-second latency budgets, for final legal determinations without human review, or as a standalone replacement for cryptographic content provenance checks. It is also unsuitable for environments where the model cannot access the raw multimodal inputs directly and must rely solely on pre-extracted text descriptions. Avoid using this prompt as the sole defense in high-stakes child safety scenarios without a mandatory human-in-the-loop review stage. The next step after implementing this prompt is to wire it into a broader safety harness with validation, retries, and logging, as described in the implementation section.

This prompt template is designed for safety engineers who need to classify unsafe content across image, audio, and video inputs within a single, unified evaluation pass. It accepts one or more modalities simultaneously and produces a structured safety verdict that includes per-modality harm classifications, cross-modal correlation flags, and an overall risk score. The template uses square-bracket placeholders for all variable inputs, making it straightforward to integrate into a production safety pipeline where inputs are assembled programmatically before inference.

textCopy
You are a multimodal safety classifier. Your task is to analyze the provided content across all supplied modalities and produce a structured safety classification.

## INPUT MODALITIES
[IMAGE_INPUT]
[AUDIO_TRANSCRIPT_INPUT]
[VIDEO_FRAME_DESCRIPTIONS_INPUT]
[VIDEO_TRANSCRIPT_INPUT]

## SAFETY POLICY
[SAFETY_POLICY_DOCUMENT]

## HARM CATEGORIES
[HARM_TAXONOMY]

## INSTRUCTIONS
1. Analyze each supplied modality independently for policy violations.
2. For each modality, identify all applicable harm categories from the taxonomy.
3. Assign a severity level (LOW, MEDIUM, HIGH, CRITICAL) to each detected harm.
4. Provide a confidence score (0.0 to 1.0) for each classification.
5. Extract specific evidence from the content that supports each classification.
6. Identify cross-modal correlations where content across modalities reinforces or contradicts safety signals.
7. Produce an overall risk score (0.0 to 1.0) and recommended action (ALLOW, FLAG, BLOCK, ESCALATE).
8. If any modality is not supplied, mark it as NOT_PROVIDED and skip its analysis.

## OUTPUT SCHEMA
[OUTPUT_SCHEMA]

## CONSTRAINTS
- Do not generate content that would violate the safety policy.
- If confidence is below [CONFIDENCE_THRESHOLD], set the recommended action to ESCALATE.
- Flag any input that appears obfuscated, encoded, or structured to evade detection.
- For CRITICAL severity findings, always recommend BLOCK or ESCALATE regardless of confidence.
- Do not output the original harmful content in evidence fields; describe it instead.

## EXAMPLES
[FEW_SHOT_EXAMPLES]

Produce the classification output now.

To adapt this template, start by replacing the modality placeholders with your actual content. [IMAGE_INPUT] should contain base64-encoded images or image descriptions depending on your model's capabilities. [AUDIO_TRANSCRIPT_INPUT] and [VIDEO_TRANSCRIPT_INPUT] should contain pre-processed transcripts. [VIDEO_FRAME_DESCRIPTIONS_INPUT] should contain sampled frame descriptions with timestamps. The [SAFETY_POLICY_DOCUMENT] placeholder must be replaced with your organization's actual policy text, not a summary. [HARM_TAXONOMY] should contain your complete category hierarchy with definitions. [OUTPUT_SCHEMA] must be a strict JSON schema that your validation layer can enforce. [CONFIDENCE_THRESHOLD] should be a numeric value between 0.0 and 1.0 that gates escalation behavior. [FEW_SHOT_EXAMPLES] should include at least one clean example, one single-modality violation, and one cross-modal correlation example to calibrate the model's behavior.

Before deploying this prompt, build a validation harness that parses the JSON output and verifies it against your [OUTPUT_SCHEMA]. The harness should reject responses with missing required fields, invalid enum values, or confidence scores outside the 0.0–1.0 range. Implement a retry loop with a maximum of two additional attempts using a repair prompt if validation fails. Log every classification decision with the full input hash, output, and validation result for audit and regression testing. For CRITICAL severity findings or ESCALATE recommendations, route the decision to a human review queue before taking enforcement action. Test the prompt against a golden dataset that includes modality-specific evasion techniques such as harmful content hidden in image backgrounds, whispered audio instructions, and video content where the visual and audio tracks carry conflicting signals.

Integrating a multimodal safety classifier into a production application requires more than a single API call. The prompt must sit inside a harness that handles input preprocessing, model routing, output validation, and decision enforcement. For this prompt, the harness accepts image, audio, or video inputs alongside optional text context, normalizes them into the model's expected format, and routes the request to a vision-capable or multimodal model. The harness then validates the structured safety classification output against the expected schema, checks confidence scores against configurable thresholds, and enforces the appropriate gating decision: block, flag, escalate, or allow.

Start by building a thin orchestration layer that wraps the prompt. The layer should: (1) validate that each input modality is supported and within size limits before calling the model; (2) inject the prompt template with the preprocessed inputs and any session-level risk context; (3) parse the model's JSON output and validate it against the classification schema, including required fields like harm_category, confidence_score, severity_level, and cross_modal_correlation_flags; (4) apply a retry policy with exponential backoff if the output fails schema validation or the model returns a malformed response—limit retries to 2–3 attempts before escalating to a human review queue; (5) log the full prompt, raw response, parsed classification, and any validation errors to an observability system for audit and debugging. For high-risk domains, always route classifications above a severity threshold or below a confidence floor to a human reviewer before any automated action is taken.

Model choice matters. Use a multimodal model with strong vision and audio understanding capabilities. If your latency budget allows, consider running a secondary verification pass with a different model for high-severity classifications to reduce false positives. For batch or high-throughput scenarios, implement a two-stage pipeline: a fast, lightweight safety classifier for initial triage, and the full multimodal prompt only for inputs that trigger an ambiguous or elevated risk score. Store all classification decisions with their evidence, timestamps, and model version identifiers. This audit trail is essential for regulatory review, false-positive analysis, and prompt iteration. Avoid wiring the classifier directly to irreversible actions—always insert a configurable gating layer that can be tuned or disabled without redeploying the prompt itself.