Voice Profile Optimization

Reduce token cost of voice profiles while preserving all constraints. Run after /interfluence analyze or on any existing profile.

Announce: "Optimizing voice profile for token efficiency."

Prerequisites

Load the profile via MCP:

1. profile_list(projectDir) to discover available voices
2. profile_get(projectDir) for the base profile
3. profile_get(projectDir, voice) for each delta

If no profile exists, tell the user to run /interfluence analyze first.

Step 1: Measure Baseline

Count the approximate token length of each profile (base + deltas). Report:

Base profile: ~X tokens (Y lines)
Delta "blog": ~X tokens (Y lines)
Total: ~X tokens

Step 2: Deduplicate Rules

Scan all H2 sections for semantically equivalent constraints. Common duplications:

• Same avoidance in both "Vocabulary & Diction" and "Anti-Patterns"
• Same pattern described in "Sentence Structure" and "Structure Patterns"
• Same directive in base and a delta (delta should only contain differences)

For each duplicate found, keep the most specific version in the most appropriate section. Remove the other.

Step 3: Cut Meta-Commentary

Remove sentences that describe the profile's own design rather than instructing the model:

• "The briefing voice inherits this but uses it structurally rather than lyrically"
• "This section captures the author's relationship with..."
• "Based on the corpus analysis, we observe that..."

These describe how the profile was made, not what the voice sounds like. Delete them.

Step 4: Convert Atmosphere to Directives

Rewrite evocative but non-actionable prose into Do/Don't form:

Before: "Cultural References: Draws from a rich well of literary fiction, particularly Banks and Le Guin, creating an atmosphere of thoughtful science-fiction discourse."

After: "Cultural References: Reference Banks, Le Guin, and literary SF when illustrating points. Don't reference pop culture, Marvel, or mainstream tech influencers."

Step 5: Trim Examples

Where both a long corpus quote AND a Do/Don't pair demonstrate the same pattern:

• Keep the Do/Don't pair (more token-efficient)
• Remove the long quote
• Preserve at least one corpus quote per section for authenticity

Step 6: Merge Overlapping Sections

If two sections have heavy overlap after deduplication:

• Anti-Patterns + AI-Tells → merge into "Banned Patterns"
• Only merge if >50% content overlaps; otherwise keep separate

Step 7: Verify Completeness

Before saving, verify:

• All 7 required H2 sections present (Overview, Sentence Structure, Vocabulary & Diction, Tone & Voice, Structure Patterns, Cultural References, Anti-Patterns/Banned Patterns)
• Every directive is actionable (do X / don't do Y)
• At least one corpus quote per section preserved
• No constraint was removed — only consolidated or reworded

Step 8: Save and Report

Save optimized profiles via profile_save. Report:

Optimization complete:
- Base: X tokens → Y tokens (Z% reduction)
- Delta "blog": X → Y (Z%)
- Total: X → Y (Z%)

Changes:
- Deduplicated N rules across sections
- Cut N meta-commentary sentences
- Converted N atmospheric descriptions to directives
- Trimmed N redundant corpus quotes
- Merged N overlapping sections

If reduction is <10%, note that the profile was already well-optimized.