Harness Prompt Feature

The prompt feature owns reusable prompt templates, dynamic system prompts, message placeholders, few-shot examples, rendered prompt caching, and injection of runtime context into prompts.

Source Inspiration

LangChain prompt primitives include string prompts, chat prompts, message templates, image prompts, few-shot prompts, structured prompts, prompt loading, and prompt unit tests:

• prompt core: https://github.com/langchain-ai/langchain/tree/master/libs/core/langchain_core/prompts
• prompt tests: https://github.com/langchain-ai/langchain/tree/master/libs/core/tests/unit_tests/prompts
• dynamic prompt middleware: https://github.com/langchain-ai/langchain/blob/master/libs/langchain_v1/langchain/agents/middleware/types.py

Responsibilities

• Render system, user, assistant, and tool message templates.
• Support message placeholders.
• Support runtime context variables.
• Support few-shot examples.
• Support multimodal prompt content.
• Support structured prompt schemas.
• Validate required variables.
• Keep rendered prompts observable.
• Cache rendered prompts when safe.
• Support dynamic prompts through middleware.

Template Rules

Prompt rendering should be explicit and typed. A template should declare:

• required variables
• optional variables and defaults
• output message role
• output content blocks
• whether rendered output is cacheable
• redaction hints for rendered content

Prompt rendering errors should happen before model invocation and should be classified separately from provider errors.

Runtime Context

Dynamic prompts may use:

• run metadata
• thread id
• user/application context
• selected model profile
• retrieved context
• memory summaries
• current date/time when explicitly configured

Dynamic prompts should not access global state implicitly. Time, random ids, and external data should come from RunContext or configured providers so tests can remain deterministic.