Prompt Clarity & Refusal Triage

Paste the prompt that got refused (or that just gave you a weak answer). This tool separates two very different problems: the request is off-limits, or the request is fine but under-specified. Only the second one is fixable by rewriting.

Your prompt

vague no schema bypass attempt harmful ask dual-use

Diagnosis

–clarity
Nothing analysed yetPaste a prompt and hit Analyse.
    
        
    
          
          
        
    
      
    
    
    

    
    
      
    
        
        
        
      
    
      
    
        
    1. The request itself
    
        
      
          
    • Operational harm: making weapons, malware, drugs, or doing targeted harm. No wording makes this fine — the problem is the goal, not the sentence.
      • 
          
    • Targeted abuse: harassment, doxxing, impersonation, non-consensual content.
      • 
        
    
        
    2. Missing legitimate context
    
        
      
          
    • Dual-use topics (security, medicine, chemistry, law) are often answered when the framing shows the real use case: authorized testing, coursework, patient education, a genuine professional need.
      • 
          
    • This is not a loophole — it's the information the model needs to give you a responsible, actually useful answer.
      • 
        
    
        
    3. Phrasing, not policy
    
        
      
          
    • Vague prompts get cautious, hedged, generic answers that feel like refusals. Often the model simply doesn't know what you want.
      • 
          
    • This is the most common cause and the one worth fixing.
      • 
        
    
      
    
      
    
        
    Say what it's for
    
        
    "I'm a [role] doing [task] for [audience]" unlocks more than any clever preamble. It is context, not a password.
    
        
    Ask for the shape of the answer
    
        
    Name the format, the length, and what to leave out. Most disappointing answers are a specification failure, not a policy failure.
    
        
    Take it to the right tool
    
        
      
          
    • Security testing → on systems you own or are contracted to test; use a lab VM or CTF platform.
      • 
          
    • Medical dosing → a pharmacist, a formulary, or the drug's own label.
      • 
          
    • Chemistry procedures → your institution's lab safety officer and published protocols.
      • 
          
    • Legal specifics → a lawyer licensed in your jurisdiction.
      • 
        
    
        
    When a model is wrong about its own rules
    
        
    Legitimate requests get refused by mistake all the time. Rephrase more plainly, restate the context, or ask the model to name the specific policy it thinks applies.
    
      
    
      
    
        
    Prefixes like "ignore all previous instructions", roleplay framings, "developer mode", "DAN", or claimed authority are extraction techniques, and they're the reason you got refused — not a route past the refusal.
    
        
      
          
    • They fail because safety behaviour is trained into the model, not bolted on as a text filter to be talked around.
      • 
          
    • They also degrade answer quality: models that "comply" with a persona request become less accurate, not more.
      • 
          
    • If your goal needs a bypass to work, the goal is the thing to change.
      • 
          
    • Studying these techniques is a legitimate research field (AI red-teaming) — but that's done on systems you're permitted to test, with results disclosed responsibly, not as a consumer tool for extracting restricted content.
      •