A better prompt can reduce unsupported answers, but no phrase can guarantee truth. Hallucination is not merely a formatting mistake. A model may generate a plausible claim, quotation, citation, calculation, or explanation that is not supported by evidence. Clear instructions improve the probability of useful behavior; they do not turn generated text into a verified source.
NIST describes confabulation as a central generative AI risk. OWASP treats misinformation and overreliance as application risks, while official model-provider guidance consistently recommends clear instructions, relevant context, explicit constraints, grounding, and iterative evaluation.
The practical goal is therefore not “make the model never hallucinate.” It is to make unsupported claims less likely, easier to identify, less consequential, and unable to pass into real decisions without verification.
1. Define the evidence boundary before asking the question
State what information the answer may use. Without an evidence boundary, the model may combine supplied material with general patterns learned during training and present the result as one seamless answer.
Useful boundaries include:
- use only the attached document
- use only the facts listed in the context
- use current official sources retrieved with an approved search tool
- answer from general knowledge but label facts that require fresh verification
- do not infer missing dates, prices, names, or quotations
For document-grounded work, tell the model what to do when the answer is absent. A strong instruction is not “be accurate.” It is “If the supplied material does not contain the answer, state that the evidence is insufficient and list what additional source is needed.”
Also define the time boundary. A question about a current plan, law, product feature, office holder, or security advisory should include the required as-of date and a tool or source that can retrieve current information.
2. Provide relevant context in a structure the model can follow
More context is not always better. Long, mixed, duplicated, or contradictory material can make the answer less reliable. Supply only the sources needed for the task and separate instructions from evidence.
A reusable structure is:
Task:
What decision or output is required?
Allowed evidence:
Which documents or tool results may be used?
Constraints:
What must not be assumed or invented?
Output:
What sections, fields, or table are required?
Uncertainty rule:
What should happen when evidence is missing or conflicting?
For long contexts, place the source material first and the exact question after it. Use clear headings or delimiters so that a quotation, web page, or customer message is not confused with an instruction.
Do not paste secrets, private records, or unapproved personal data merely to improve context. Use the AI Tool Privacy and Security Checklist Before Adoption before supplying real organizational material.
3. Separate facts, inferences, recommendations, and unknowns
Ask the model to classify statements instead of blending them.
A useful answer format can require four fields:
| Field | Meaning |
|---|---|
| Supported fact | Directly stated or measured in the allowed evidence |
| Inference | A conclusion derived from facts, with assumptions shown |
| Recommendation | A judgment based on stated criteria and tradeoffs |
| Unknown | Information that cannot be established from the available evidence |
This structure exposes where confidence comes from. It also prevents a recommendation from looking like a measured fact.
For consequential work, request a claim ledger containing the exact claim, supporting source, applicable scope, checked date, and verification status. The model may help construct the ledger, but a person or deterministic process must verify important entries.
Use calibrated language. “The document states” is different from “This suggests,” and both are different from “We could not verify.”
4. Ask for citations and quotations in a verifiable form
Requesting citations can improve traceability, but generated citations can also be wrong. A citation is useful only when a reviewer can open it and confirm that it supports the claim.
Ask for:
- source title and publisher
- direct URL or stable identifier
- exact section, page, table, or timestamp
- short quotation when permitted and necessary
- the claim supported by that source
- date accessed for volatile information
- conditions such as region, plan, version, or sample
Do not ask the model to “add some sources” after writing an answer. That encourages decorative references. Require each important claim to be attached to evidence during generation.
Then verify the citation separately. Check that the source exists, the metadata matches, the quoted text is present, and the surrounding context supports the conclusion. Use How to Verify AI Answers, Sources, and Factual Claims for the full process.
5. Constrain the answer to an explicit schema
Open-ended prose makes omissions and unsupported transitions harder to notice. For repeated work, define a schema that requires the model to expose evidence and gaps.
For example:
Return a table with these columns:
Claim | Evidence | Scope | Status | Missing information
Allowed status values:
verified | conditional | conflicting | unverified
Do not leave Evidence or Status blank.
Do not convert unverified claims into recommendations.
A schema does not prove correctness, but it makes validation easier. Software can reject missing fields, unexpected status values, invalid dates, unsafe URLs, or outputs that do not match the required structure.
Prefer positive instructions that describe the desired behavior. “Use only supported facts and mark missing evidence” is often clearer than a long list of prohibited mistakes.
6. Use retrieval, calculation, and code tools for the right claims
A model should not be the sole source for current facts or exact computation when a more reliable tool is available.
Use approved retrieval or grounding for:
- current prices and product capabilities
- recent events and office holders
- changing policies and regulations
- release notes and security advisories
- exact quotations from a document
Use a calculator or code execution for:
- arithmetic and unit conversion
- aggregation and counting
- date calculations
- reproducible transformations
- checking a formula against original inputs
Tool output is still not automatically correct. Search can retrieve an outdated page, code can use the wrong inputs, and a model can misread a result. Record the query, source, inputs, output, and transformation so the result can be repeated.
The prompt should tell the model which tool to use and what to do if the tool is unavailable. It should not silently replace missing retrieval with memory.
7. Ask adversarial follow-up questions before accepting the result
A single polished answer should not end the process. Use follow-up prompts to expose weak assumptions and missing evidence.
Examples include:
- Which claims are not directly supported by the supplied sources?
- What conditions would make this conclusion false?
- Which dates, prices, names, or quotations require fresh verification?
- Are any sources secondary summaries of the same original material?
- What evidence contradicts the recommendation?
- Which part of the answer depends most on an assumption?
- What would a skeptical domain expert challenge first?
Do not treat the model's self-critique as independent verification. It may identify useful issues, but it is still generated by the same class of system. Important claims require original evidence, deterministic checks, reproduction, or qualified human review.
For high-impact workflows, route the proposed result through the approval design in Designing AI Workflows That Require Human Approval.
8. Build an evaluation set for repeated prompts
A prompt that works once is not a production control. Save representative inputs and expected properties, then test the prompt whenever the model, tool, source format, or instruction changes.
Include:
- ordinary cases
- missing evidence
- contradictory sources
- outdated information
- ambiguous wording
- long context
- irrelevant or malicious instructions inside documents
- requests that should be refused or escalated
- calculations with known answers
- examples where “unknown” is the correct result
Evaluate more than writing quality. Measure unsupported claim rate, citation validity, omission rate, correct uncertainty, format compliance, correction time, and human acceptance.
A change that produces more fluent answers but hides uncertainty is a regression. A shorter answer that clearly separates verified facts from unknowns may be more useful.
Prompt template for evidence-bound answers
You are assisting with a factual research task.
Task:
[State the exact question and decision.]
Allowed evidence:
[List supplied documents or approved retrieval tools.]
Rules:
- Use only the allowed evidence for factual claims.
- Separate supported facts, inferences, recommendations, and unknowns.
- Cite the exact source section for each important claim.
- State the applicable date, region, plan, version, or sample.
- If evidence is missing or conflicting, say so explicitly.
- Do not invent quotations, citations, numbers, dates, or product behavior.
Output:
1. Direct answer
2. Claim and evidence table
3. Uncertainty and conflicting evidence
4. Verification actions still required
This template is a starting point. Adapt it to the task, model, tools, and risk level, then test it with representative cases.
Prompt patterns that create false confidence
- “Give me the definitive answer” when the evidence is incomplete.
- “Do not say you are uncertain” or “never refuse.”
- Asking for sources after the conclusion is already written.
- Supplying many unlabelled documents with conflicting dates.
- Treating a role prompt such as “You are an expert” as proof of expertise.
- Requesting exact numbers without providing data or a calculation tool.
- Asking the model to verify its own answer without opening original sources.
- Requiring a recommendation even when mandatory facts are unknown.
- Assuming a low temperature guarantees factual accuracy.
Short hallucination-reduction checklist
- Is the allowed evidence and as-of date explicit?
- Is the context relevant, structured, and free of unapproved sensitive data?
- Must the answer separate facts, inferences, recommendations, and unknowns?
- Are citations tied to specific claims and source sections?
- Is the output schema easy to validate?
- Are current facts and calculations delegated to appropriate tools?
- Have contradictory evidence and failure conditions been requested?
- Is there an evaluation set for repeated use?
- Will important claims receive independent verification and human approval?
Sources reviewed
- NIST Generative AI Profile (opens in a new window)
- Google prompt design strategies (opens in a new window)
- Anthropic prompt engineering best practices (opens in a new window)
- OpenAI Model Spec: Express uncertainty (opens in a new window)
- OWASP LLM09:2025 Misinformation (opens in a new window)
Sources checked: 2026-07-20