Your voice AI prompt,
professionally audited.

FULL AUDIT
      ----------
      # Voice AI Prompt Audit: Bland Car Rentals

        Total Score: 46 / 80
        Rating: Needs Significant Improvement

      ## 1. QUICK SUMMARY

      This prompt has strong foundational instincts — the Midwestern tone, backchannel logic, and customer empathy are thoughtful touches. However, **the prompt is significantly over-engineered and internally contradictory**, which is likely causing the agent to behave unpredictably and contributing to caller hang-ups. The cognitive load on the model is immense, with competing instructions that will create hesitation, repetition, and awkward phrasing in live calls.

      ---

      ## 2. SCORECARD
        --------
        Voice-First Language: 5/10
        Persona Clarity & Consistency: 6/10
        Conversational Flow & Turn-Taking: 4/10
        Handling Real-World Call Behavior: 6/10
        Conciseness & Signal-to-Noise: 3/10
        Human-Sounding Behavior: 6/10
        Edge Cases & Failure Modes: 5/10
        Business Outcome Alignment: 5/10

      
      ## 3. WHAT'S WORKING

      **Tone and persona guidance is genuinely thoughtful:**
      > "amiable concise cheerful apologetic... midwestern charm... 'no worries, take your time!'... Basically, 'Minnesota nice,' with a sprinkle of brevity."

      This is excellent — it gives the model a real personality to embody, not just a job function.

      **Empathy-first error handling:**
      > "Instead of 'I think you gave me a partial number' we'd want to say 'Any chance you could please double check that number, I may have missed it but it sounds a few digits short?'"

        ## 4. WHAT NEEDS IMPROVEMENT

        ### A. **Contradictory Instructions Create Decision Paralysis**

        The prompt repeatedly contradicts itself, which will cause the model to freeze or behave inconsistently:

        > "Don't repeat yourself... Please don't repeat yourself..."

        But also:
        > "You then need to read back their Rental Record Number... Repeat the number back to them using the below format..."

        And:
        > "if a customer corrects a small mistake you should just start back at that number"
        ## 5. TOP 3 PRIORITIZED IMPROVEMENTS

        ### 1. **Simplify and Consolidate the Core Flow**

        Rewrite the prompt as a clean 3-step process:
        1. Greet warmly and ask why they're calling
        2. Request rental record or confirmation number
        3. Read it back using paced speech

        Remove all the conditional backchannel math and digit-counting logic. Instead, give a simple rule:
        > "If the caller is still reading digits, say 'mhmm' or 'got it' and wait. Once they've finished, read the number back slowly."

        ### 2. **Fix the Opening to Reduce Hang-Ups**
        ## FINAL THOUGHT

        You clearly understand voice AI at a deep level — the backchannel logic, the pacing tokens, the Midwestern charm. But this prompt is trying to do too much. The model is drowning in rules, and callers are experiencing hesitation, odd phrasing, or robotic behavior as a result.

        **The fix isn't more instructions — it's fewer, clearer ones.** Strip it down to the essential flow, trust the model to handle variations, and focus your explicit guidance on the moments that matter most: the greeting, the number readback, and recovering from