Prompt Development Guide

Nexent packages built-in prompt YAML under sdk/nexent/core/prompts/. Backend services collect business data and select templates. The SDK loads resources and renders fields with supplied parameters.

Files and interfaces

• Localized resources live under zh/ and en/, then in purpose-specific folders such as agent, meta, evaluation, and tool. Both languages have matching relative paths. The meta/ folder contains prompts used to generate prompts, Agent drafts, and skill drafts.
• Agent resources live in zh/agent/ and en/agent/; manager and worker roles use agent_manager.yaml and agent_worker.yaml. Chat-title generation uses agent/generate_chat_title.yaml. Role templates have system_sections, role-specific stage fields, and final_answer.
• Human interaction, context summary, and answer verification live in each language's agent/human_interaction.yaml, agent/context_summary.yaml, and agent/answer_verifier.yaml. The Chinese context-summary and verifier resources are simple translations of the English originals.
• load_prompt(language, relative_path) loads an independent mapping relative to the language directory, for example load_prompt("zh", "agent/human_interaction"). The .yaml suffix is optional. Use render_prompt_text(source, parameters) to render Jinja text with runtime values.
• Backend services call the SDK using the resource language, purpose, and filename. User-defined prompt generation templates remain in the database; their built-in defaults come from SDK YAML.

Parameters

Callers pass current values to the SDK renderer. Agent configuration or conversation state supplies values such as duty, constraint, few_shots, and delegated tasks. NL2Agent, skill creation, and other workflows supply their tool names, skill drafts, document content, and related values. Missing required Jinja parameters raise an error instead of sending unresolved template placeholders to the model.

Extending a template

1. Add YAML to the appropriate language and purpose folder, for example zh/evaluation/judge.yaml. Preserve the workflow's field names and placeholders.
2. Use the resource language and relative path at the call site. Existing Agent templates also pass SDK bundle validation. Meta prompts use meta/nl2agent.yaml and meta/nl2skill.yaml.
3. Add parameter rendering and package resource tests. SDK pyproject.toml declares YAML package data in the two-level subdirectories.
4. Run the consuming Backend service tests to verify final message content, language selection, and user template fallback.