Prompt Development Guide

This guide explains how Nexent prompt templates are organized under backend/prompts/ and how to extend them for new agents.

📂 File Layout & Naming

• Core templates live in backend/prompts/ using {agent_type}_agent.yaml or {scope}_prompt_template.yaml.
• Utility templates are under backend/prompts/utils/ for meta generation (e.g., prompt/title helpers).

🧩 Template Structure

Each YAML may contain:

• system_prompt: role, responsibilities, execution flow, tool/sub-agent usage rules, Python code constraints, and examples.
• planning: initial_facts, initial_plan, update hooks before/after facts or plans.
• managed_agent: prompts for delegating tasks and collecting reports from sub-agents.
• final_answer: pre/post messages to shape final output.
• tools_requirement: priorities and guardrails for tool usage.
• few_shots: examples to steer behavior.

🔄 Variables

Common placeholders for runtime rendering:

• tools, managed_agents
• task, remaining_steps
• authorized_imports
• facts_update, answer_facts

📑 Key Templates

• Manager agents: manager_system_prompt_template.yaml, manager_system_prompt_template_en.yaml
• Managed agents: managed_system_prompt_template.yaml, managed_system_prompt_template_en.yaml
• Knowledge summary: knowledge_summary_agent.yaml, knowledge_summary_agent_en.yaml
• File analysis: analyze_file.yaml, analyze_file_en.yaml
• Cluster summary: cluster_summary_agent.yaml, cluster_summary_reduce.yaml (and _zh variants)
• Utilities (utils/): prompt_generate*.yaml, generate_title*.yaml

🚀 How to Extend

1. Copy the closest existing template and adjust system_prompt/planning for your scenario.
2. Keep placeholders intact unless intentionally removed.
3. Align tool lists with actual tools available to the agent; update authorized_imports if needed.
4. Validate with a small task to ensure flows (Think → Code → Observe → Repeat) produce the expected behavior.

✅ Standards & Tips

• Use executable code fences for runnable snippets: py, and display-only fences for non-executable examples.
• Prefer keyword args for tool calls; avoid excessive tool invocations per step.
• Keep comments and docstrings in English and respect repository coding rules.