Multi-Tool Result Aggregation and Synthesis Prompt Template

This prompt is designed for agent infrastructure engineers who are orchestrating multiple tool calls in parallel and need a single, coherent, and trustworthy summary for the end user. The core job-to-be-done is to fuse disparate, potentially conflicting structured outputs from APIs, databases, or internal services into one narrative that resolves contradictions, eliminates redundancy, and faithfully cites the origin of each fact. The ideal user is building a production AI system where a user asks one question, the system fans out to several tools, and the raw JSON blobs must be synthesized before they reach the UI.

You should use this prompt when you have two or more tool results that address the same user intent and you need to produce a single, non-redundant answer. It is particularly valuable when the tools may return overlapping or contradictory information—for example, one inventory API showing an item in stock and a shipping API showing it backordered. The prompt template is built to force the model to surface and resolve these conflicts rather than silently picking the first result. It also requires the model to cite which tool provided each claim, making the output auditable. This is not a prompt for simple RAG or single-source summarization; it is for multi-source synthesis where provenance matters.

Do not use this prompt when you are only interpreting a single tool's output, when the tools return completely unrelated data, or when the user needs raw data instead of a summary. It is also the wrong choice if your primary goal is to decide whether to make a follow-up tool call—that requires a separate routing or planning prompt. If your system operates in a regulated domain where every claim must be traceable to a specific record, you must pair this prompt with a post-processing validation harness that checks citation integrity before the response reaches the user. The prompt alone reduces hallucination risk but does not eliminate it; always run eval checks for cross-tool contradiction handling and unattributed claims in production.

This template is designed for agent infrastructure engineers orchestrating parallel tool calls. Its primary job is to consume the raw outputs of multiple tools, resolve any conflicts between them, remove redundant information, and produce a single, coherent summary for the user. Every factual claim in the summary must be explicitly linked back to the specific tool that provided it. Use this prompt when you need a reliable, auditable synthesis step before presenting information to a user or passing it to another agent. Do not use this template if the tool outputs are already in a final user-ready format or if the task is a simple concatenation rather than a true synthesis.

markdownCopy
You are an expert synthesis agent. Your task is to process the outputs of multiple tools, resolve any conflicts, and produce a single, coherent summary for the user.

## TOOL OUTPUTS
Below are the results from the tools that were called. Each result is wrapped in an XML tag indicating the tool's name.

[TOOL_OUTPUTS]

## USER QUERY
The user's original request that triggered these tool calls is provided for context.

[USER_QUERY]

## INSTRUCTIONS
1.  **Analyze:** Review all tool outputs. Identify the core facts, key data points, and any direct answers to the user's query.
2.  **Resolve Conflicts:** If two or more tools provide contradictory information, you must:
    *   Explicitly state the conflict in your response.
    *   Do not arbitrarily choose one source over another.
    *   If possible, suggest a reason for the conflict (e.g., different data freshness, different scopes) and recommend a follow-up action to resolve it.
3.  **Remove Redundancy:** If multiple tools report the same fact, mention it only once in the final summary.
4.  **Synthesize:** Weave the resolved, non-redundant information into a single, flowing narrative that directly answers the user's query. Structure the response logically, using bullet points or sections if the information is complex.
5.  **Cite Meticulously:** Every factual statement in your summary must be followed by a citation tag indicating its source tool. Use the format `[source: tool_name]`. If a statement synthesizes information from multiple tools, list all relevant sources, e.g., `[sources: tool_a, tool_b]`.
6.  **Handle Gaps:** If the user's query cannot be fully answered by the provided tool outputs, clearly state what information is missing. Do not invent or infer missing details.

## OUTPUT FORMAT
Produce your response in two parts:

**Part 1: Synthesis Summary**
A direct, comprehensive answer to the user's query, following all instructions above.

**Part 2: Processing Notes**
A brief, bulleted list of any conflicts found, data gaps identified, or assumptions made during the synthesis process.

To adapt this template, replace the [TOOL_OUTPUTS] placeholder with the raw string outputs from your parallel tool calls. It is critical to wrap each tool's output in a uniquely identifying XML tag (e.g., <tool_a>...</tool_a>) before insertion. The [USER_QUERY] placeholder should contain the exact original user prompt. For high-stakes applications, add a [RISK_LEVEL] placeholder and a corresponding instruction to escalate to a human reviewer if the risk level is 'high' and unresolvable conflicts are detected. You can also add a [CONSTRAINTS] section to enforce rules like 'Do not mention internal tool names to the user' or 'Summarize in under 150 words'.

Before deploying, test this prompt with a golden dataset of multi-tool outputs that includes known conflicts, redundant data, and incomplete answers. Your evaluation should check for citation accuracy, conflict transparency, and hallucination rate. A common failure mode is the model resolving a conflict by silently choosing the data from the first or most authoritative-sounding tool, which defeats the purpose of the audit trail. Another is omitting citations for synthesized or rephrased statements. Implement a post-processing validation step that scans the final summary for any factual claim not immediately followed by a [source: ...] tag, and flag it for review or automatic retry.

This prompt is designed to sit between a tool orchestration layer and the user-facing response. It receives a structured payload containing the original user query and the raw outputs from multiple parallel tool calls. The application is responsible for collecting all tool results, handling timeouts and partial failures, and assembling the input object before invoking the model. Do not pass raw tool outputs directly to the model without first wrapping them in the expected input schema, which should include a tool_call_id, tool_name, status (success, error, timeout, empty), and the raw output for each tool call.

Validation and retry logic should be implemented in the application layer, not in the prompt. After the model returns a synthesized response, validate that every factual claim in the output can be traced to at least one tool result in the input. A lightweight eval harness can check for unattributed claims by extracting all cited tool names from the response and comparing them against the provided tool list. If the model fabricates a tool name or makes a claim without a citation, trigger a retry with a stronger instruction to only use information from the provided tool outputs. Set a maximum of two retries before falling back to a templated message that lists raw tool results without synthesis.

Logging and audit trails are critical for debugging cross-tool contradictions. Store the full input payload, the model's synthesized response, and the validation result in a structured log. Include a contradiction_flag field set by a secondary eval prompt that checks whether the synthesis acknowledges or resolves conflicts between tool outputs. For high-stakes domains, route responses with unresolved contradictions to a human review queue before they reach the end user. Avoid using this prompt for real-time safety-critical decisions without a human-in-the-loop gate.