Expert Needed to Fix Prompt / System Prompt Collision in Virtual Agent Harness

# Expert Needed to Fix Prompt / System Prompt Collision in Virtual Agent Harness We have developed a virtual agent harness designed to prevent hallucinations and ensure the agent responds correctly during live user conversations. We are currently seeing what appears to be a collision between the **system prompt**, the **production prompt**, and possibly parts of the agent harness / code. The result is that the virtual agent does not always respond correctly when engaged by a user. In some cases, the agent appears to prioritize the wrong instruction layer, repeat the wrong question, ignore direct user questions, or move into the wrong part of the workflow. We need someone with deep experience in: * AI agent prompt architecture * System prompt vs production prompt separation * Agent harness design * LLM instruction precedence * Voice or chat agent workflow debugging * State management in agent conversations * Preventing hallucinations and prompt conflicts * Reviewing code, prompts, transcripts, and agent behavior together ## Project Scope We will provide: * The agent code * The system prompt * The production prompt * Call transcripts * Our analysis of what is going wrong * Examples of the correct behavior expected * Access details or files needed to review the setup The freelancer will need to: 1. Review the full setup. 2. Identify where the system prompt, production prompt, and/or agent harness are conflicting. 3. Fix the issue in the code and/or prompt structure. 4. Clearly separate responsibilities between: * System prompt * Production prompt * Harness / workflow logic * Runtime state handling 5. Ensure the agent responds correctly to users during live interaction. 6. Provide the corrected files and a clear explanation of what was changed. ## Important Working Process This is **not an open-ended hourly project**. We are looking for a freelancer who can review the materials, understand the issue, and provide a fixed-price solution. After you make the fix, we will test it on our development server. We will provide a video of the call, the transcript, and additional analysis if the issue is not fully resolved. We expect to iterate until the issue is fixed. ## Payment Terms This must be structured as a **fixed-price milestone**. Payment will only be released after: * The issue is fixed * The agent is tested on our development server * The agent behaves correctly in the test calls * The prompt collision / instruction precedence issue is resolved * The agent no longer repeats incorrect questions or ignores the user’s direct request * We confirm the fix through transcript and/or call testing We will not pay for an open-ended review that does not result in a working fix. You may inspect the materials first and then provide a hard fixed price for the fix, but the paid milestone must be tied to successful resolution of the issue. ## Current Suspected Issue We believe this is a prompt precedence and separation problem. The system prompt and production prompt may both be trying to control the same behavior. We need a clean separation where: * The system prompt controls identity, safety rules, guardrails, and high-level boundaries. * The production prompt controls the live task, product-specific context, and conversation behavior. * The harness/code controls state, workflow progression, completed steps, and validation. * The agent does not allow one layer to override another incorrectly. ## Examples of Problems We Are Seeing The agent may: * Ask the same question repeatedly * Ignore a user’s direct question * Jump into product identification too early * Continue a workflow step after it has already been answered * Treat general questions as troubleshooting questions * Fail to respect completed conversation state * Respond from the wrong prompt layer * Follow production prompt instructions that conflict with system-level behavior ## Desired Outcome The final result should be a stable agent setup where: * The virtual agent responds naturally and correctly * Prompt responsibilities are clearly separated * The harness prevents hallucinations without creating workflow collisions * User questions are answered appropriately * The agent follows the correct workflow only when appropriate * Completed steps are remembered and not repeated * Direct user questions are handled before workflow progression * The system is ready for further live testing ## Ideal Freelancer You should have proven experience fixing AI agent behavior issues involving: * Prompt collisions * System vs developer vs production prompt hierarchy * LLM workflow orchestration * Conversational state management * Voice agent or customer support agent flows * RAG or hallucination-prevention harnesses * Debugging real transcripts against expected behavior Please do not apply if you only do basic prompt writing. This requires someone who can review the full architecture and fix the behavior at the prompt/code/workflow level. ##