Stale Content Detection Harness Prompt

This prompt is an engineered component for AI quality engineers and platform teams who need to automate content freshness monitoring at scale. It is designed for production systems that run scheduled audits, CI/CD pipelines for knowledge base maintenance, or guardrails that prevent outdated information from reaching end users. The ideal user is someone building a monitoring harness—not someone performing a one-off manual review. You should use this prompt when you have a content item with associated metadata (publication date, last-updated timestamp, document type) and a domain-specific freshness policy that defines acceptable staleness windows. The prompt expects these structured inputs and returns a staleness report with severity levels, recommended actions, and confidence scores that can be consumed by downstream alerting or routing systems.

This is not a prompt for ad-hoc freshness checks, casual content reviews, or workflows where a human is expected to read the output and make a subjective call without further automation. If you need a quick staleness opinion on a single document, a simpler classification prompt will suffice. The harness prompt is overengineered for that use case. It is also not a replacement for a full retrieval pipeline or a real-time data validation system—it evaluates one content item at a time against a declared policy. Do not use this prompt when the freshness policy itself is undefined or when the content lacks reliable date metadata. The prompt's confidence scores degrade sharply when input timestamps are missing, ambiguous, or contradictory, and it will flag this uncertainty rather than guess.

Before integrating this prompt, ensure you have defined your domain-specific freshness thresholds, have a mechanism for extracting or providing publication dates, and have a plan for what happens when the prompt returns a high-severity staleness alert. The output is structured for machine consumption—expect to wire it into alerting systems, dashboard pipelines, or automated content refresh queues. If your workflow requires human review of staleness decisions, build that review step into the harness layer, not into the prompt itself. The prompt is a detection component; the surrounding application owns the escalation path.

This template is the core instruction set for your stale content detection harness. It accepts a content item, a set of freshness rules, and an optional evidence context, then produces a structured staleness report. The prompt is designed to be wired into a CI/CD pipeline, a scheduled audit job, or an on-demand content review workflow. Every square-bracket placeholder must be replaced with your application's data before the prompt is sent to the model. The template enforces a strict JSON output schema so downstream validators can parse the result without ambiguity.

textCopy
You are a content freshness auditor. Your task is to evaluate whether a piece of content is stale, suggest a replacement if it is, and generate an alert with a severity level and confidence score.

## INPUT

**Content to evaluate:**
[CONTENT]

**Content metadata:**
- Publication date: [PUBLICATION_DATE]
- Content type: [CONTENT_TYPE]
- Domain: [DOMAIN]
- Authoritative source: [IS_AUTHORITATIVE_SOURCE]

**Freshness rules:**
[FRESHNESS_RULES]

**Current date:** [CURRENT_DATE]

**Optional evidence context (retrieved passages that may supersede this content):**
[EVIDENCE_CONTEXT]

## OUTPUT SCHEMA

Return a single JSON object with these fields:

{
  "content_id": "string, the identifier provided in metadata or a generated hash",
  "is_stale": "boolean, true if the content exceeds freshness thresholds or has been superseded",
  "staleness_severity": "string, one of: 'critical', 'high', 'medium', 'low', 'none'",
  "confidence_score": "number between 0.0 and 1.0, how confident you are in the staleness determination",
  "staleness_reasons": ["array of strings, specific reasons why the content is or is not stale"],
  "recommended_action": "string, one of: 'replace', 'review', 'keep', 'deprecate', 'append_notice'",
  "replacement_suggestion": "string or null, a specific suggestion for what should replace this content, citing evidence if available",
  "replacement_evidence": ["array of strings or null, specific passages from EVIDENCE_CONTEXT that support the replacement"],
  "temporal_gap": "string or null, human-readable description of the time gap between content publication and current date or superseding evidence",
  "alert_message": "string or null, a concise alert message suitable for logging or notification systems, only present if is_stale is true"
}

## CONSTRAINTS

1. If [EVIDENCE_CONTEXT] is empty or not provided, base staleness solely on [FRESHNESS_RULES] and the gap between [PUBLICATION_DATE] and [CURRENT_DATE].
2. If [EVIDENCE_CONTEXT] contains newer information that directly contradicts or supersedes [CONTENT], mark is_stale as true regardless of the publication date gap.
3. Do not fabricate replacement evidence. Only cite passages that actually appear in [EVIDENCE_CONTEXT].
4. If you are uncertain about staleness, set confidence_score below 0.7 and set recommended_action to 'review'.
5. For domains marked as high-risk in [FRESHNESS_RULES], default to 'review' when confidence_score is below 0.85.
6. Never output markdown fences or extra text outside the JSON object.
7. If [CONTENT] is empty or unreadable, return is_stale as null and set recommended_action to 'review' with an alert_message explaining the input was invalid.

After pasting this template into your prompt layer, replace each bracketed variable with data from your content store, freshness policy engine, and retrieval pipeline. The [FRESHNESS_RULES] placeholder should contain domain-specific thresholds such as 'financial earnings data expires after 90 days' or 'breaking news content expires after 24 hours.' The [EVIDENCE_CONTEXT] placeholder should be populated with passages from your retrieval system that may supersede the content under review. If your application does not have a retrieval pipeline, pass an empty string and the prompt will fall back to date-gap analysis. Before deploying, run this prompt against a golden dataset of known-stale and known-fresh content to calibrate your severity thresholds and confidence score expectations.

The Stale Content Detection Harness Prompt is designed to be called programmatically, not used as a one-off chat interaction. It expects a structured input payload containing the content to evaluate, a domain context, and optional freshness windows. The output is a machine-readable JSON report with severity levels, recommended actions, and confidence scores. This structured contract makes it suitable for integration into automated content monitoring pipelines, pre-publication review gates, and scheduled knowledge base audits.

To wire this prompt into an application, wrap it in a function that accepts a StalenessCheckRequest object with fields: content_text (string), content_type (enum: article, report, documentation, data_snapshot, market_data), publication_date (ISO 8601 string or null), domain (enum: news, finance, legal, medical, technical, general), and freshness_window_days (integer or null). The function should inject these into the prompt's [INPUT] placeholder, call the model with response_format set to json_object (or equivalent structured output mode), and validate the returned JSON against a schema that requires: staleness_score (0-1 float), severity (enum: fresh, aging, stale, critical), findings (array of objects with claim, evidence_age, recommended_action, and confidence), and overall_recommendation (string). If the model returns malformed JSON, implement a retry loop with the repair prompt from the Output Repair and Validation Prompts pillar, passing the raw output and the expected schema. Log every evaluation with the input hash, model version, timestamp, and staleness score for trend analysis.

For CI/CD integration, invoke this harness as a gate in your content pipeline before publication or after scheduled content refreshes. If severity returns critical or stale, block the pipeline and route to a human review queue. For scheduled knowledge base audits, run the harness against all documents older than their domain-specific freshness window, aggregate results into a StalenessAuditReport, and generate prioritized refresh tickets. Avoid using this prompt for real-time streaming data freshness checks—those require latency monitoring, not content analysis. For high-stakes domains like finance and healthcare, always require human review of stale and critical findings before taking automated action. Pair this harness with the Temporal Relevance Scoring Prompt for pre-retrieval filtering and the Recency-Boosted Reranking Prompt for post-retrieval ordering to build a complete freshness-aware RAG pipeline.