Profanity in Creative Writing Allowlist Prompt Template

This prompt is designed for content platforms and creative writing tools that need to enforce a profanity policy without destroying artistic expression. The core job-to-be-done is context-aware classification: distinguishing profanity used as a literary device (character voice, dialogue, poetic effect) from profanity used to abuse, harass, or violate platform conduct standards. The ideal user is a product engineer or trust-and-safety engineer integrating an AI moderation layer into a system that accepts user-generated fiction, screenplays, poetry, or narrative game content. You need this prompt when your existing keyword-based or simple classifier is generating an unacceptable rate of false positives on legitimate creative work, causing user complaints, support tickets, and creator churn.

You should not use this prompt if your platform has a zero-tolerance profanity policy with no exceptions for creative context. In that case, a simpler keyword-blocking or regex-based approach is more appropriate and cheaper to operate. Similarly, do not use this prompt as your primary defense against harassment, hate speech, or threats—those require dedicated safety classifiers. This prompt is a calibration layer that sits between broad content filters and final publishing decisions. It assumes that some profanity is permitted in context, and its value is in reducing over-refusal without opening the door to genuinely abusive content. The prompt requires the full text of the creative work as input, along with your platform's specific profanity policy definitions and any category taxonomy you use for content decisions.

Before deploying this prompt, confirm that you have a clear, documented profanity policy that distinguishes between allowable creative use and disallowed abusive use. The prompt cannot invent your policy boundaries—it can only apply them. You should also prepare a set of golden test cases covering edge cases: dialogue with period-appropriate slurs in historical fiction, profanity-laden character rants that are not directed at another user, poetry with explicit language, and abusive messages disguised as creative writing. Run these through the prompt before production and measure both false positive rate (legitimate creative work flagged) and false negative rate (abusive content allowed through). If either rate is unacceptable, tune the policy language in the prompt rather than adding more rules. The next section provides the copy-ready template you can adapt.

This prompt template is designed to be copied directly into your prompt management system, IDE, or orchestration layer. It accepts a passage of creative writing and returns a structured classification that distinguishes artistic, character-driven, or literary profanity from abusive, harassing, or gratuitous profanity. The template uses square-bracket placeholders that you must replace with your specific policy definitions, examples, and output schema before deployment. Do not use this prompt as-is without adapting the placeholders to your platform's content policy and risk tolerance.

textCopy
You are a content classification system for a creative writing platform. Your task is to evaluate whether profanity in a submitted passage serves a legitimate artistic, narrative, or character-development purpose, or whether it constitutes abusive, harassing, or gratuitous profanity that violates platform policy.

## INPUT
[PASSAGE]

## CONTEXT
- Author intent: [AUTHOR_INTENT_CONTEXT]
- Genre: [GENRE]
- Target audience: [AUDIENCE]
- Surrounding narrative context: [NARRATIVE_CONTEXT]

## CLASSIFICATION CRITERIA
Evaluate the passage against the following allowlist conditions. Profanity is PERMITTED when it:
1. Serves character voice, dialect, or authentic dialogue
2. Advances narrative tone, setting, or emotional stakes
3. Functions as literary device (e.g., juxtaposition, rhythm, thematic emphasis)
4. Appears in quoted speech within a larger narrative framework
5. Is consistent with the established genre conventions (e.g., literary fiction, screenplay, poetry)

Profanity is DISALLOWED when it:
1. Targets a specific person or group with abusive or harassing intent
2. Serves no narrative, character, or artistic purpose (gratuitous use)
3. Constitutes hate speech, threats, or incitement to violence
4. Appears in authorial voice directed at the reader abusively
5. Violates [PLATFORM_SPECIFIC_POLICY_REFERENCE]

## EXAMPLES
[ALLOWLIST_EXAMPLES]
[DISALLOWLIST_EXAMPLES]

## OUTPUT SCHEMA
Return a JSON object with the following fields:
{
  "classification": "allowed" | "disallowed" | "review",
  "confidence": 0.0-1.0,
  "profanity_instances": [
    {
      "term": "string",
      "position": "approximate location in passage",
      "purpose": "character_voice" | "narrative_tone" | "literary_device" | "dialogue" | "gratuitous" | "abusive" | "harassment" | "hate_speech",
      "justification": "brief explanation of classification"
    }
  ],
  "overall_justification": "summary of why the passage was classified as allowed, disallowed, or flagged for review",
  "policy_references": ["specific policy sections triggered"],
  "recommended_action": "publish" | "block" | "flag_for_human_review"
}

## CONSTRAINTS
- Do not refuse to classify passages that contain profanity; this system exists specifically to evaluate creative writing
- When confidence is below [CONFIDENCE_THRESHOLD], set classification to "review" and recommended_action to "flag_for_human_review"
- If the passage contains both allowed and disallowed profanity, classify the overall passage as "disallowed" and flag individual instances separately
- Do not modify, censor, or rewrite the passage; only classify it
- For poetry and experimental forms, apply a wider interpretation of literary purpose

## RISK LEVEL
[RISK_LEVEL]

Adaptation guidance: Replace [PLATFORM_SPECIFIC_POLICY_REFERENCE] with a link or summary of your actual content policy. Populate [ALLOWLIST_EXAMPLES] and [DISALLOWLIST_EXAMPLES] with 3-5 real passages from your platform that represent clear boundary cases. Set [CONFIDENCE_THRESHOLD] based on your risk tolerance—start at 0.85 and adjust after evaluating false positive and false negative rates on a golden test set. The [RISK_LEVEL] placeholder should contain any escalation rules, such as "If classification is 'disallowed' and confidence is above 0.9, auto-block; otherwise flag for human review." Test this prompt against the screenplay excerpts, novel passages, and poetry samples in your eval suite before deploying to production. If your platform serves minors, add an additional constraint requiring human review for any passage with profanity in young adult or children's literature categories.

This prompt is designed to operate as a classification microservice within a broader content moderation pipeline. It should be invoked after initial ingestion but before publishing, editing, or flagging workflows. The model receives a creative writing excerpt and returns a structured classification with a decision (allow, flag, escalate), a rationale, and extracted profanity instances. The application layer should treat this output as a scoring signal, not a final publish/discard gate. The prompt expects the model to distinguish between literary profanity (character voice, dialogue, poetic device) and abusive or harassing profanity, but the final decision should combine this classification with user trust level, content context, and platform policy.

Integration pattern: Wrap the prompt in an async function that accepts [CONTENT_EXCERPT], [CONTENT_TYPE] (e.g., screenplay, novel_chapter, poem), and [AUTHOR_CONTEXT] (e.g., established_author, new_user, workshop_participant). The function should: (1) assemble the prompt with these variables, (2) call the model with response_format set to the JSON schema defined in the prompt, (3) validate the output against the schema, (4) apply platform-specific rules. For validation, check that decision is one of the three allowed values, rationale is non-empty, and profanity_instances is an array where each entry has a word, context, and classification. If validation fails, retry once with the error message appended to the prompt. If the second attempt fails, route to a human review queue with the original excerpt and both failed outputs attached.

Threshold design: Use the classification output to drive routing logic. allow decisions with confidence above 0.9 can proceed automatically for trusted authors. flag decisions should always queue for human review, but the reviewer should see the model's rationale and extracted instances pre-highlighted to accelerate their decision. escalate decisions should block publication and notify a trust-and-safety specialist immediately. For model choice, a mid-tier model (GPT-4o-mini, Claude Haiku) is sufficient for most excerpts under 2,000 words. For longer works, chunk the content by scene or chapter and aggregate classifications, flagging the work if any chunk returns flag or escalate. Log every classification with the model version, prompt hash, input excerpt ID, and output decision for audit trails and policy regression testing.

Failure modes to instrument: Watch for (1) false negatives where abusive profanity directed at a character is misclassified as dialogue because the model over-indexes on quotation marks, (2) false positives on period-appropriate language in historical fiction where modern profanity filters are too aggressive, (3) boundary blurring in literary fiction where profanity serves both character development and shock value simultaneously. Build an eval set from your platform's actual content decisions, including edge cases that previously caused moderator disagreement. Run this eval set against every prompt or model change. If your platform operates in regulated markets or serves minors, add a mandatory human review step for any content within 0.2 of the flag threshold, even if the model returns allow.