Meta-Prompting Harness
Internal exploration.
Objective
See whether a model writing its own instructions beats a hand-tuned prompt on a fixed task set.
Process
Built a small harness with 30 graded tasks, ran hand-written prompts against generated ones, and logged every failure with its category.
Prompt
Write the prompt that writes the prompt: given a task spec, emit a constrained instruction set.
Result
Generated prompts won on coverage and lost on precision. The hybrid — generated draft, human edit — beat both.
Learning
Keep the evaluation harness; it was worth more than any individual prompt it produced.