Claude Long-Context Evidence Assembly Prompt Template

This prompt template is designed for RAG and document intelligence teams who need Claude to synthesize an answer from a large set of retrieved documents. It is built for scenarios where the evidence set exceeds what a human could easily hold in working memory, where source documents may conflict, and where citation accuracy is non-negotiable. The core job-to-be-done is transforming a noisy, redundant, and potentially contradictory set of retrieved chunks into a single, coherent, and fully grounded answer. The ideal user is an AI engineer or product developer who has already built a retrieval pipeline and now needs a reliable assembly layer that can handle the full complexity of a 200K context window without losing fidelity or hallucinating sources.

Use this prompt when you are packing a significant portion of Claude's context window with evidence chunks and need the model to prioritize salient information, resolve contradictions explicitly, and ground every claim in a specific source. This template assumes your retrieval system may return noise, near-duplicates, and conflicting data, and it includes explicit instructions for handling each of these failure modes. It is appropriate for regulated or high-trust domains where an unsupported claim can have serious consequences, as it enforces a strict citation discipline and requires the model to abstain when evidence is insufficient. The prompt also includes a token budget allocation strategy, ensuring that instructions, evidence, and the output itself all fit within the model's limits without silent truncation.

Do not use this prompt for simple Q&A over a single document or for workflows where the retrieval step already returns a perfectly ordered, non-redundant set of facts. In those cases, a lighter-weight RAG prompt without conflict resolution and deduplication instructions will be faster and cheaper. Similarly, avoid this template if your application requires real-time streaming answers where the full evidence assembly step would introduce unacceptable latency. If your use case involves multi-turn conversation rather than a single-shot research task, you will need to pair this with a conversation state management pattern. Before implementing, confirm that your retrieval pipeline is returning chunks with stable, unique identifiers that can be used in citations, as the prompt's grounding guarantees depend on traceable source references.

This template is designed for RAG and document intelligence teams using Claude's extended context window to synthesize answers from large sets of retrieved evidence. It enforces strict citation requirements, handles conflicting sources explicitly, and includes token budget allocation instructions to prevent the model from ignoring late-context documents. Replace every square-bracket placeholder with runtime values before inference. The template assumes you have already retrieved, deduplicated, and ordered your evidence chunks before assembly—this prompt governs how Claude uses that evidence, not how you retrieve it.

textCopy
You are an evidence-based research assistant. Your task is to answer the user's question using ONLY the provided evidence documents. You must cite your sources inline and flag any conflicts between documents.

## EVIDENCE DOCUMENTS
Below are [EVIDENCE_COUNT] documents retrieved for this query. Each document has a unique citation ID in square brackets. Use these IDs for all citations.

[DOCUMENTS]

## USER QUESTION
[USER_QUESTION]

## INSTRUCTIONS
1. Read all evidence documents before drafting your answer. Do not skip documents near the end of the context window.
2. Answer the question using ONLY information present in the evidence documents. If the evidence is insufficient to answer fully, explicitly state what is missing rather than speculating.
3. For every factual claim in your answer, include an inline citation using the format [Doc_X] where X is the document's citation ID. Place citations immediately after the claim they support.
4. If two or more documents contain conflicting information on the same point, you MUST:
   a. State the conflict explicitly.
   b. Present each conflicting position with its supporting citation.
   c. Do not choose a winner unless one source is clearly more authoritative and you can explain why.
5. Structure your answer as follows:
   a. **Direct Answer**: A concise answer to the question in 1-3 sentences.
   b. **Detailed Explanation**: Full explanation with inline citations.
   c. **Evidence Gaps**: Any aspects of the question the evidence could not address.
   d. **Source Conflicts**: Any contradictions found between documents, with citations.
6. Do not fabricate citations. If you reference a document, the citation ID must match one provided in the evidence.
7. If the evidence contains information that appears outdated, note the document's date (if available) but still report what the document states.
8. Prioritize documents marked as [HIGH_SALIENCE] when resolving ambiguity, but still report conflicting lower-salience evidence.

## OUTPUT FORMAT
Return your response in the following structure:

DIRECT ANSWER:
[Your concise answer here]

DETAILED EXPLANATION:
[Your full explanation with inline citations]

EVIDENCE GAPS:
[What the evidence could not address]

SOURCE CONFLICTS:
[Any contradictions found, with citations for each position]

## CONSTRAINTS
- Do not use knowledge outside the provided evidence documents.
- Do not omit citations for any factual claim.
- Maximum response length: [MAX_OUTPUT_TOKENS] tokens.
- If the user question appears to request harmful, illegal, or policy-violating content, respond with: "I cannot assist with this request based on the provided evidence and my usage policies."

Adaptation notes: Replace [EVIDENCE_COUNT] with the number of documents you're providing. The [DOCUMENTS] placeholder should be replaced with your pre-formatted evidence chunks, each prefixed with a citation ID like [Doc_1], [Doc_2], etc. Include document metadata (date, source type, authority level) inside each chunk if available. The [HIGH_SALIENCE] marker should be applied during your retrieval ranking step—only mark documents your reranker scored above your salience threshold. Set [MAX_OUTPUT_TOKENS] based on your application's latency budget and expected answer length. For regulated use cases, add a human review step before the output reaches end users. Test this template with conflicting documents and insufficient evidence to verify the model follows the abstention and conflict-reporting instructions before deploying.

This prompt is designed to be the final assembly step in a RAG pipeline, not a standalone chat interaction. The application layer is responsible for retrieval, deduplication, and token budgeting before the prompt is constructed. The prompt template receives pre-ranked evidence, a user query, and output constraints, then produces a synthesized answer with inline citations. The harness must validate that the model's output respects the citation contract, stays within the evidence, and explicitly abstains when the evidence is insufficient.

Pre-Assembly Pipeline: Before calling the model, the application must: (1) Retrieve and deduplicate passages from your vector store or search index. (2) Rank passages by salience using a cross-encoder or LLM-based reranker. (3) Allocate token budget across instructions, evidence, and output space—reserve at least 20% of the context window for the model's response. (4) Truncate low-ranking passages to fit the budget. (5) Assign stable source IDs to each passage for citation matching. Post-Output Validation: After receiving the model response, run a structured validation harness: parse the output for citation markers (e.g., [1], [2]), verify every cited source ID exists in the provided evidence set, flag any unsupported factual claims using an NLI model or a secondary LLM judge, and check for the abstention keyword (e.g., 'INSUFFICIENT_EVIDENCE') when no relevant passages were provided. If validation fails, increment a retry counter and re-invoke the model with the validation errors appended as [CONSTRAINTS] feedback. Stop retrying after 2 attempts and escalate to a human review queue.

Observability and Logging: Log the full assembled prompt, the model's raw response, the validation results, and any retry attempts as a single trace. Include a trace ID, prompt version, model ID, token counts (input and output), and the source document IDs used. This trace data is essential for debugging citation failures, detecting evidence drift, and auditing outputs in regulated workflows. Model Choice: This template is tuned for Claude 3.5 Sonnet and Opus, which handle long-context instruction following and inline citation reliably. If using a different model, validate citation adherence and abstention behavior with a golden test set before production deployment. When to Escalate: If the model consistently hallucinates citations, ignores the abstention instruction, or exceeds the allocated output token budget, do not add more prompt instructions. Instead, shorten the evidence set, increase the retry budget, or switch to a model with stronger long-context instruction following.