prompt-optimization

You can improve LLM accuracy on structured tasks by applying prompt repetition during the prefill stage. This targets non-reasoning models to enhance reliability for parsing and unit tests without adding latency to the final response.

The best way to increase Haiku agent accuracy for unit tests and linting is prompt repetition. Configuring environment flags to repeat prompts boosts accuracy by 4-5x on structured tasks like automated QA checks.

You configure prompt repetition for Haiku agents by setting environment variables. Enable the feature with LOKI_PROMPT_REPETITION=true and define the repetition count with LOKI_PROMPT_REPETITION_COUNT before running your task.

Prompt repetition is specifically designed for non-reasoning LLMs like Haiku agents. It targets structured tasks such as list operations and parsing, where repeating prompts ensures consistent results without relying on advanced reasoning.

Prompt repetition does not add latency because it operates in the prefill stage. By handling repeated prompts during this initial processing phase, the overall response time remains unaffected while improving accuracy.