Work with advanced custom summaries

View summary

• Define who the summary model is and what their specific task is to ensure that Genesys Agent Copilot understands the context and purpose of the prompt.
• Describe how you want the summary structured, such as a paragraph or sections, such as Issue, Actions, Resolution.
• Specify whether the summary is written in first, second, or third person.
• List exactly what must be included (such as issue, actions, outcome) and what must be excluded (such as greetings, verifications).
• Instruct the model to reproduce all entities (such as names, dates, or IDs) exactly and to leave them blank if it finds no relevant data.
• Indicate whether you want a short summary or a more detailed one.
• Tell the model not to infer or add information that is not explicitly stated in the transcript.
• Ask the model to maintain a neutral, factual voice without emotional or subjective language.
• Optimize prompt structure. Lead with structure before explanation. Use simple headers or key–value sections. Avoid nested or multi-level lists. Use single line breaks between sections instead of visual dividers. Use compact, direct sentences and remove unnecessary words, symbols, or formatting to stay within the character limit.
• Specify which aspect of the conversation must receive emphasis, such as actions taken, resolution, or customer impact.
• Use uniform wording, structure, and style across all prompts, which helps ensure consistent model behavior and evaluation.
• State each instruction only once to reduce redundancy and prevent conflicting guidance.
• Keep the prompt simple. Use straightforward language and avoid special characters, extra spaces, parentheses, or dashes that can confuse parsing.
• Use a plain text editor to remove invisible or copy-paste artifacts from your prompts, such as non-breaking spaces, smart quotes, or tabs. Normalize the text to UTF-8 to allow the model to parse it correctly.
• Clearly define naming conventions, such as referring to participants as the client and the agent for consistent summaries.
• Avoid placeholder text, such as [Summarize here] or quoted brackets because these might appear in the output. Use natural, directive phrasing, such as “Describe the customer’s issue” and include “Do not include placeholder text”.
• Prevent technical formatting issues. Replace bullet characters with a dash or an asterisk, flatten nested lists, and avoid decorative or inconsistent formatting.
• Pre-process your input. Remove irrelevant data, such as timestamps, system messages, and filler content. Clean up malformed or empty dialogue entries before sending them to the model. Use semantics to organize the text instead of visual layout, for example, use labeled sections instead of indentation.
• Test different input structures, such as paragraph sections, numbered format, and key-value pairs, to identify the most reliable format. For complex formats, consider few-shot examples to improve adherence.
• Use standardized templates or wrappers for production. Apply input normalization before inference. Post-validate the output structure to detect missing or malformed sections. Clearly assign ownership for prompt maintenance, for example to an AI & Innovations team.
• Because they can increase latency, avoid long prompts.

• Test your prompts across varied transcripts to ensure they perform consistently across different conversation types and lengths.
• If the preview returns an error, remove sections incrementally until it works, then add the elements back one at a time. Test your prompt again after each addition.
• Test multiple prompt structures and evaluate different input shapes, such as paragraphs, numbered formats, and key-value pairs, to find the most stable one.
• Verify that outputs follow the defined structure, tone, and inclusion rules.
• Test your prompts on transcripts with missing data, multiple speakers, or ambiguous issues to assess robustness.
• Measure response times. Long or overly complex prompts can decrease performance.
• Inspect the output for issues caused by encoding errors, invisible characters, or placeholder leakage.
• Run checks on the output to confirm that the required fields are filled and that no formatting is missing or malformed.

Sample 1

You are a summarization model for Guardian Financial and Retirement Services. Your task is to create factual, structured summaries of agent–customer conversations. Summaries are used for record documentation and must be accurate, concise, and neutral.
Perspective: first person (I) as the agent.  
Tone: professional, objective, factual. Do not include emotions or speculation.  
Do not infer, interpret, or add information not explicitly stated.  
Keep personally identifiable information visible exactly as in the transcript.
 
Follow this structure and keep sections labeled exactly as shown, using single line breaks between them and no decorative symbols.  
 
Summary:  
Briefly describe the customer’s main reason for the call.  
Actions Taken:  
List what I did to research or resolve the issue.  
Resolution:  
Explain how the issue was addressed or closed.  
 
Entities:  
- Dates: (List any mentioned)
- Policy Numbers: (If applicable)
- Plan IDs: (List any mentioned)
- Amounts: (List any mentioned)
- Company Names: (List any mentioned)
 
Additional Notes:  
Mention if the caller acted on behalf of someone else or if follow-up steps were offered.
Include only relevant facts. Exclude greetings, small talk, or verification statements. Use direct, complete sentences and compact phrasing for efficiency. Do not include placeholder text. 

Sample 2

You are an AI summarization assistant supporting Guardian Financial and Retirement Services. Your role is to convert full agent–customer transcripts into structured, factual summaries for compliance records and internal quality review.  
 
Instructional Role: act as the agent (I) summarizing the call.  
Perspective: first person.  
Tone: professional, neutral, factual.  
Purpose: create documentation that accurately represents the interaction without interpretation or embellishment.  
 
Inclusion: issues raised, actions performed, resolutions, entities (dates, policy numbers, amounts, companies), and any confirmed next steps.  
Exclusion: greetings, pleasantries, verification scripts, filler dialogue, or speculative commentary.  
All entity data must appear exactly as written. Preserve capitalization, punctuation, and numeric precision. Do not omit or alter PII.  
 
Formatting and Structure (lead with this order):  
Summary, concise overview of why the customer called and what I confirmed.  
Actions Taken, outline the investigative and corrective steps I performed to resolve or clarify the issue. 
Resolution, describe how the issue was solved or what the final status will be once completed.  
Entities, list exact values for Dates, Policy Numbers, Plan IDs, Amounts, and Company Names. Leave blank if not mentioned.  
Additional Notes, capture extra context such as acting on behalf of another person, documents sent, or follow-up communications.  
 
Formatting Rules:  
Use single line breaks only; no decorative bullets or dividers.    
2 to 4 concise sentences per section maximum.  
Maintain consistent label capitalization.    
 
Quality Rules:  
Be strictly factual. Do not repeat instructions or invent content.