Citation Style Enforcement Prompt

This prompt is for RAG system builders and AI engineers who need every generated answer to follow a specific citation style, not just contain citations. It enforces formatting rules for author names, dates, titles, URLs, and access dates at the output layer. Use it when your product must produce bibliographically consistent references for academic publishing, legal filings, regulatory submissions, or enterprise knowledge bases where style guide compliance is auditable. The core job-to-be-done is transforming a set of retrieved documents into an answer where every reference is formatted according to a target style guide like APA 7th, MLA 9th, Chicago 17th, or a custom house style, and where deviation from that style is treated as a critical output defect.

The ideal user is an AI engineer integrating a RAG pipeline into a product where the output is consumed by a downstream system or a professional audience that expects formal citation formatting. Required context includes the target style guide specification, the raw source metadata for each retrieved document (author, title, date, URL, publisher, access date), and the generated answer text that needs citation formatting. Do not use this prompt if you only need inline source IDs or bracketed reference numbers; those workflows are better served by the Inline Citation Formatting System Prompt or Deterministic Citation Output Contract Prompt in this pillar. This prompt is also not appropriate when source metadata is incomplete or unreliable, as the model will hallucinate missing fields to satisfy the style guide, creating a false sense of bibliographic accuracy.

Before deploying this prompt, ensure your retrieval pipeline captures the metadata fields required by your target style guide. A common failure mode is feeding the prompt with documents that have titles but no publication dates, leading the model to invent 'n.d.' placeholders or guess years. Implement a pre-generation metadata completeness check that flags documents with missing required fields and either routes them for human review or excludes them from citation. For high-stakes domains like legal or regulatory submissions, pair this prompt with a post-generation citation verification step that parses each formatted reference and validates it against the original source metadata before the answer reaches the user.

This prompt template enforces consistent citation formatting across every answer your RAG system generates. Instead of hoping the model follows your style guide, you give it explicit rules for inline citations, reference lists, quote handling, and source metadata inclusion. The template supports APA, MLA, Chicago, or custom styles by replacing the [CITATION_STYLE_RULES] placeholder with your specific formatting requirements. Use this when your answers must be auditable, when downstream systems parse citations programmatically, or when multiple team members contribute to the same knowledge base and need uniform output.

textCopy
You are a citation-disciplined answer generator. Your job is to produce answers grounded in the provided source documents, with every factual claim backed by a properly formatted citation.

## CITATION STYLE RULES
[CITATION_STYLE_RULES]

## SOURCE DOCUMENTS
[SOURCE_DOCUMENTS]

## USER QUESTION
[USER_QUESTION]

## OUTPUT FORMAT
[OUTPUT_SCHEMA]

## CONSTRAINTS
- Every factual claim, statistic, date, name, or specific assertion must include an inline citation pointing to the exact source document and section.
- Verbatim quotes must be wrapped in quotation marks and must not exceed [MAX_QUOTE_LENGTH] words or [MAX_QUOTE_RATIO]% of the source passage.
- When sources conflict, surface the disagreement explicitly with citations to each conflicting source rather than choosing one silently.
- If the provided sources cannot answer the question, state that clearly instead of fabricating an answer. Use the phrase: [ABSTENTION_PHRASE]
- Do not cite sources outside the provided [SOURCE_DOCUMENTS] list.
- Include [SOURCE_METADATA_FIELDS] in every citation.
- Format the reference list at the end according to [REFERENCE_LIST_FORMAT].

## EXAMPLES
[FEW_SHOT_EXAMPLES]

## RISK LEVEL
[RISK_LEVEL]

To adapt this template, start by replacing [CITATION_STYLE_RULES] with your specific formatting instructions. For APA, specify author-date parenthetical citations and a references section with hanging indents. For a custom internal style, define exactly how source IDs, page numbers, and document titles appear inline. The [OUTPUT_SCHEMA] placeholder should contain your expected response structure—plain text with bracketed citations, a JSON object with a claims array and citations map, or Markdown with footnote-style links. Replace [FEW_SHOT_EXAMPLES] with at least two examples showing correct citation behavior and one showing correct abstention when sources are insufficient. Set [RISK_LEVEL] to high if answers feed into compliance, legal, or clinical workflows; this triggers stricter adherence to the abstention rule and requires human review before deployment. Before shipping, run this prompt through your eval suite with a golden dataset that includes source-conflict cases, missing-evidence cases, and edge cases where quote limits would be exceeded. Measure citation precision (are citations pointing to the right source?), citation recall (are all claims cited?), and format compliance (do citations match the style guide exactly?). If format compliance drops below your threshold, add more few-shot examples or tighten the [CONSTRAINTS] section rather than lengthening the style rules.

The Citation Style Enforcement Prompt is not a standalone generator; it is a post-processing step that sits between your RAG answer synthesis and the final output delivered to the user or downstream system. In a production pipeline, the model first generates a grounded answer with inline citations using your standard RAG prompt. That answer, along with the target style guide specification, is then passed to this enforcement prompt to normalize every citation into the required format. This separation of concerns—answer generation vs. citation formatting—keeps your primary RAG prompt focused on evidence synthesis and reduces the risk of format drift contaminating answer quality.

To wire this into an application, build a two-stage pipeline. Stage one calls your primary RAG model to produce an answer with raw citations. Stage two calls the Citation Style Enforcement Prompt with the raw answer as [RAW_ANSWER] and the target style as [CITATION_STYLE]. The output must pass through a citation format validator before release. For APA, validate that in-text citations match (Author, Year) patterns and that the reference list is alphabetized. For MLA, check for (Author Page) parentheticals and a Works Cited header. For custom styles, define a regex or JSON schema validator that matches your required format. If validation fails, retry the enforcement prompt once with the validation error appended to the input as additional context. If the retry also fails, log the raw answer and the failed output, then fall back to the raw answer with a warning flag rather than blocking the response entirely.

Observability is critical because citation format errors are often the first signal of a deeper problem—model version drift, context window truncation dropping reference details, or retrieval returning documents with inconsistent metadata. Log the following fields per request: citation_style, input_citation_count, output_citation_count, validation_pass (boolean), retry_count, and format_violations (list of specific errors). Set up an alert if the validation failure rate exceeds 5% over a rolling hour, as this may indicate a model behavior change or a new document format breaking your citation extraction. For high-stakes domains like legal or clinical RAG, route all validation failures to a human review queue rather than auto-retrying. The reviewer can confirm whether the citation content is correct even if the format is wrong, preventing a formatting error from blocking a medically or legally accurate answer.

Model choice matters here. This prompt works well with instruction-following models like gpt-4o, claude-3-5-sonnet, or gemini-1.5-pro. Avoid using smaller models (below ~7B parameters) for strict style enforcement unless you have fine-tuned them on your specific citation format, as they tend to hallucinate author names or invent page numbers when the raw answer lacks them. If you must use a smaller model for cost reasons, add a citation completeness check after enforcement: verify that every citation in the output corresponds to a citation in the raw input. Any added citations are likely hallucinations and should trigger a fallback to the raw answer. Finally, never pass user-facing text through this prompt without first confirming that the raw answer itself passed your grounding and hallucination checks—formatting a wrong answer beautifully is worse than leaving it unformatted.