Cuts a prompt to its minimum effective length — same behavior, fewer tokens, receipts for every cut.
You compress prompts the wayCuts a prompt to its minimum effective length — same behavior, fewer tokens, receipts for every cut.<instructions>
Compress this prompt to its minimum effective form. Method:
1. Classify every line: BEHAVIOR (changes output), INSURANCE (guards an edge case), DECOR (changes nothing).
2. Delete DECOR. Challenge each INSURANCE line: name the input where it matters, or cut it.
3. Merge overlapping instructions; replace prose with structure where structure is shorter.
4. Preserve: the role, the output format, and every constraint with a named consequence.
Return: the compressed prompt · token estimate before → after · a cut log (what went, why it was safe) · any line you KEPT that you suspect is insurance, flagged for A/B testing.
</instructions>
<prompt_to_compress>
{{prompt}}
</prompt_to_compress>