Grounded Answer Guardrails
Use this to design prompts that reduce unsupported claims and invented details.
Use case: A researcher or workplace team can make document-based AI answers more traceable and safer to verify.
Best tools
- NotebookLM — Grounds responses in user-provided sources and exposes supporting source locations.
- Claude — Follows detailed evidence constraints across long source material.
Fill in your details
Answer the questions below. Your answers are placed into the prompt automatically.
Your filled prompt (preview)
You are my Grounded Answer Guardrails designer. Design prompt rules that reduce unsupported claims and invented details.
========================================
PERSONAL CONTEXT (filled by me)
========================================
Task the AI will perform: [TASK_TYPE]
Sources the AI may use: [ALLOWED_SOURCES]
Behaviours to forbid: [FORBIDDEN_BEHAVIOURS]
How evidence must be shown: [CITATION_RULE]
How to mark unknowns: [UNCERTAINTY_RULE]
Audience for answers: [AUDIENCE]
Risk if the AI invents facts: [RISK_LEVEL]
Sample user query to protect: [SAMPLE_QUERY]
If any field is blank or still shows brackets, list the gap, continue with labelled assumptions, and never invent unverified facts, numbers, citations, or quotes.
========================================
WORKFLOW YOU MUST FOLLOW
========================================
STEP 1 — Threat model
List how the AI might invent or overclaim on [TASK_TYPE] given [RISK_LEVEL].
STEP 2 — Guardrail pack
Write enforceable rules for sources, refusals, citations, and uncertainty using my fields.
STEP 3 — Ready prompt
Produce a ready-to-paste system/user prompt pair that embeds the guardrails.
STEP 4 — Test cases
Give pass/fail test queries including [SAMPLE_QUERY].
========================================
OUTPUT FORMAT
========================================
Use these exact headings:
A. Invention risks
B. Guardrail rules
C. Ready-to-paste prompt
D. Citation behaviour
E. Uncertainty behaviour
F. Pass/fail tests
G. Escape hatches to avoid
H. Review checklist
Rules: prefer refusal or labelled uncertainty over fiction; honour [ALLOWED_SOURCES] and [FORBIDDEN_BEHAVIOURS].
You are my Grounded Answer Guardrails designer. Design prompt rules that reduce unsupported claims and invented details. ======================================== PERSONAL CONTEXT (filled by me) ======================================== Task the AI will perform: [TASK_TYPE] Sources the AI may use: [ALLOWED_SOURCES] Behaviours to forbid: [FORBIDDEN_BEHAVIOURS] How evidence must be shown: [CITATION_RULE] How to mark unknowns: [UNCERTAINTY_RULE] Audience for answers: [AUDIENCE] Risk if the AI invents facts: [RISK_LEVEL] Sample user query to protect: [SAMPLE_QUERY] If any field is blank or still shows brackets, list the gap, continue with labelled assumptions, and never invent unverified facts, numbers, citations, or quotes. ======================================== WORKFLOW YOU MUST FOLLOW ======================================== STEP 1 — Threat model List how the AI might invent or overclaim on [TASK_TYPE] given [RISK_LEVEL]. STEP 2 — Guardrail pack Write enforceable rules for sources, refusals, citations, and uncertainty using my fields. STEP 3 — Ready prompt Produce a ready-to-paste system/user prompt pair that embeds the guardrails. STEP 4 — Test cases Give pass/fail test queries including [SAMPLE_QUERY]. ======================================== OUTPUT FORMAT ======================================== Use these exact headings: A. Invention risks B. Guardrail rules C. Ready-to-paste prompt D. Citation behaviour E. Uncertainty behaviour F. Pass/fail tests G. Escape hatches to avoid H. Review checklist Rules: prefer refusal or labelled uncertainty over fiction; honour [ALLOWED_SOURCES] and [FORBIDDEN_BEHAVIOURS].
Pro tip: Require an abstention phrase and source locations, then test the prompt with an unanswerable question.
Expected output: A hallucination-resistant prompt with evidence rules and adversarial tests.