prompt-optimization

Enable prompt repetition, which duplicates the prompt so bidirectional attention in the prefill stage improves comprehension. Set LOKI_PROMPT_REPETITION=true and optionally LOKI_PROMPT_REPETITION_COUNT=2 or 3 for position-critical tasks.

Prompt repetition is a technique from arXiv 2512.14982v1 that repeats the input prompt, letting non-reasoning models attend to instructions bidirectionally during prefill. It improves accuracy on tasks like testing, linting, and parsing without increasing latency.

No, prompt repetition does not benefit reasoning models like Opus or Sonnet. The technique only improves non-reasoning LLMs such as Haiku, so the skill explicitly does not apply to reasoning agents.

Prompt repetition adds zero latency because the duplicated text is processed in the parallelizable prefill stage rather than during generation. Accuracy improvements of 36-62% are reported across linting, parsing, and test tasks.

Use 3x repetition for position-critical tasks where the model must track specific locations in the input, by setting LOKI_PROMPT_REPETITION_COUNT=3. The default 2x repetition is sufficient for most structured tasks.