Confidence-Bounded Answer Generation Prompt

This prompt is for product teams who need AI-generated answers that self-limit based on certainty. Instead of hedging or fabricating details, the model is instructed to only include claims that meet a configurable confidence threshold, explicitly mark uncertain portions, and omit claims below the threshold entirely. The core job-to-be-done is to produce a structured output that downstream systems or human reviewers can trust, where the cost of a wrong answer is higher than the cost of an incomplete answer. This is a critical distinction: the prompt optimizes for safety and auditability over fluency or completeness.

Use this prompt when your application cannot tolerate silent hallucination and transparency about uncertainty is a product requirement. Ideal contexts include clinical summarization, legal research, financial analysis, and any user-facing feature where an incorrect claim carries significant risk. The prompt requires a clear [CONFIDENCE_THRESHOLD] to be set, typically between 0.7 and 0.95, and works best when paired with a structured [OUTPUT_SCHEMA] that enforces a separation between high-confidence claims, low-confidence claims, and omitted information. It is designed for single-turn answer generation, not multi-step agentic reasoning, and assumes the model has sufficient context to assess its own certainty.

Do not use this prompt when the primary goal is creative generation, exhaustive brainstorming, or when the model's confidence assessment is known to be poorly calibrated for your domain. It is also a poor fit for low-latency, high-throughput systems where the additional reasoning tokens required for self-assessment would violate latency budgets. Before deploying, you must validate the prompt's behavior with a calibration eval set to ensure the model's self-reported confidence actually correlates with correctness. Without this validation step, the confidence scores are a false safety signal. If your use case involves regulated decisions about individuals, this prompt is a component of a human-in-the-loop system, not a replacement for human judgment.

This template is the core instruction set for a Confidence-Bounded Answer Generation system. It forces the model to introspect on the certainty of each claim it makes, rather than producing a single holistic confidence score. The prompt is designed to be dropped into an application harness that parses the structured output, enforces the threshold, and routes low-confidence claims for human review. Before using it, you must define your [CONFIDENCE_THRESHOLD], your [OUTPUT_SCHEMA], and the [CONTEXT] the model is allowed to use. This prompt is not suitable for creative writing, open-ended brainstorming, or tasks where factual precision is not the primary goal.

textCopy
You are an AI assistant that answers questions with strict confidence bounds. Your primary directive is to be accurate, not comprehensive. You will generate an answer where every factual claim is individually assessed for certainty.

## INPUT
[QUERY]

## CONTEXT (Use only this information to form your answer)
[CONTEXT]

## INSTRUCTIONS
1. Analyze the [QUERY] and determine the factual claims required to answer it.
2. For each claim, assess your confidence on a scale from 0.0 (no confidence) to 1.0 (absolute certainty) based *only* on the provided [CONTEXT].
3. Assemble your final answer using the following rules:
   - **Include** claims with a confidence score >= [CONFIDENCE_THRESHOLD].
   - **Exclude** claims with a confidence score < [CONFIDENCE_THRESHOLD]. Do not mention, allude to, or hedge these claims.
   - If excluding a claim creates a logical gap in the answer, explicitly state "Information regarding [topic] is omitted due to insufficient confidence."
4. Structure your entire response as a single JSON object conforming to the schema below.

## OUTPUT SCHEMA
[OUTPUT_SCHEMA]

## CONSTRAINTS
- Do not use information outside the provided [CONTEXT].
- Do not invent confidence scores. If you cannot determine a score, default to 0.0.
- The `answer_summary` field must be a coherent, human-readable text block built only from high-confidence claims.

To adapt this template, start by replacing the [OUTPUT_SCHEMA] placeholder with a concrete JSON schema that your application parser expects. A typical schema includes a high_confidence_answer string, an array of claims objects each with claim_text and confidence_score fields, and an array of omitted_topics strings. The [CONFIDENCE_THRESHOLD] is a critical business logic lever; a threshold of 0.9 produces highly precise but sparse answers, while 0.7 is more conversational but riskier. Wire this prompt into an application harness that validates the JSON output against the schema, checks that no claim below the threshold appears in the high_confidence_answer, and routes any output with a high proportion of omitted_topics to a human review queue. For high-risk domains, always log the full structured output before displaying the summary to the user.

The Confidence-Bounded Answer Generation Prompt is designed to be a self-limiting component within a larger AI pipeline. It should not be deployed as a standalone, user-facing endpoint without a validation harness. The prompt's core contract—that it will only return claims meeting a specified confidence threshold—must be verified programmatically before any output reaches a user or downstream system. Treat the model's self-reported confidence as a signal to be checked, not a guarantee to be trusted. The implementation harness is responsible for enforcing the threshold, validating the output schema, and deciding what happens when the model fails to comply.

Wire this prompt into your application by first defining a strict output schema. The prompt expects a [CONFIDENCE_THRESHOLD] variable (e.g., 0.85) and should return a JSON object with an array of claims, each containing a statement, a confidence_score between 0 and 1, and a grounding reference. Your harness must parse this JSON and perform a post-generation validation loop: (1) check that all returned confidence_score values are >= the threshold, (2) reject any output where the model includes claims below the threshold, and (3) verify that omitted claims are not present. If validation fails, implement a retry strategy with a maximum of 2 attempts, feeding the validation error message back into the prompt's [CONSTRAINTS] field. Log every validation failure and retry attempt for observability. For high-stakes domains like healthcare or finance, route any output that fails validation after retries to a human review queue using the Structured Uncertainty Packaging Prompt for Human Review from the same pillar.

Model choice matters significantly for this prompt. Use models with strong instruction-following and structured output capabilities, such as GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro. Avoid smaller or older models that may ignore the threshold instruction or hallucinate confidence scores. For latency-sensitive applications, consider using a smaller, faster model for the initial generation and a larger model as a judge to verify claim-level confidence accuracy. Implement a circuit breaker: if more than 5% of responses in a rolling window fail validation, stop routing traffic to the prompt and escalate to the engineering team. Never expose raw, unvalidated confidence scores to end users; always map them to clear, user-facing certainty language using the Uncertainty Disclosure Prompt for AI Outputs sibling prompt. The next step is to build a golden dataset of known-answer test cases to measure your threshold adherence rate and confidence calibration error before production rollout.