Prompt compression: reduce prompt size by 10-20% without quality regression

[apenab]

Problem

Current prompts are 300+ lines long and include repeated guidance across sections (especially overlap between strategy instructions and critical rules). This increases token usage and parsing latency.

Proposal

Implement a prompt compression pass targeting a 10-20% reduction in prompt size while preserving behavior:

• Merge redundant guidance between strategy and critical rules.
• Simplify page-selection guidance (currently overly verbose).
• Keep all non-negotiable constraints explicit and testable.

Expected Impact

• Token reduction: ~15-20%
• Lower prompt parsing latency in LLM inference.
• Minimal behavior change if compressed carefully.

Risk

• Low quality risk if constraints are preserved and validated.

Acceptance Criteria

• Prompt token count reduced by at least 10% (target 15-20%).
• No statistically significant regression on benchmark quality metrics.
• Updated prompt docs include a mapping from old sections to compressed sections.
• A/B comparison results are documented (before vs after token count, latency, quality).

Notes

Please include at least one automated or scriptable check to prevent prompt bloat regressions over time.