style-specification-building

Extract key findings with confidence levels from each analysis report, translate each finding into a constraint using ranges rather than exact values, resolve conflicts between analyses, then prioritize constraints into tiers. The output includes both a numeric profile and prose instructions.

Research on LLM prompt engineering shows models follow 5-12 well-prioritized constraints reliably, with adherence dropping beyond roughly 20. Aim for 15-25 total constraints organized into tiers, with 5-8 core MUST-follow rules in Tier 1.

At least 3 completed upstream analyses covering at least 3 different dimensions are required. Fewer than 3 analyses means insufficient data, and the specification should not be built. Specifications from 3 analyses are labeled minimal and preliminary.

First check whether the analyses measure the same feature; if not, both constraints can coexist. If they conflict on the same feature, the higher-confidence analysis sets the hard constraint and the lower-confidence one becomes a soft preference. Document both findings and the resolution rationale.

No. Function-word rates, punctuation frequencies, and other unconscious features should remain in the numeric profile only. LLMs cannot consciously control these rates, so numeric targets for unconscious features serve post-hoc validation rather than real-time instruction.

No. This specification process is designed for a single-author voice derived from one person's writing corpus. It is not designed for multi-author or composite brand voices, and merging analyses from different authors is explicitly out of scope.