Voice Validator Agent Skill

Voice Validator Skill

Overview

This skill operates a rigorous critique-and-rewrite enforcement loop for voice fidelity. It scans content against voice-specific negative prompt checklists, documents violations with evidence, fixes them while preserving intent, and rescans to confirm the revision passes — up to 3 iterations maximum.

The workflow implements the Iterative Refinement pattern: scan → document violations → revise → rescan. This ensures voice violations are caught systematically and fixed methodically without over-engineering or changing meaning.

CRITICAL CONSTRAINT: Never revise content without first scanning against the full checklist. Every violation must cite a specific quote. After 3 failed iterations, output with flagged concerns rather than continuing indefinitely.

Reference Loading Table

| Signal | Load These Files | Why |

|---|---|---|

| validating narrative content of 500+ words | narrative-patterns.md | Supplies the full narrative-fidelity rubric. |

Instructions

Phase 1: IDENTIFY TARGET

Goal: Determine the voice, mode, and content to validate.

Step 1: Identify voice target

• Determine target voice from context or user instruction
• Identify mode if applicable — casual modes may have additional specific checks
• Reference the target voice's checklist (contact user if unclear)

Step 2: Load content

• Read the content to validate
• Note content length — longer content is more prone to drift

Gate: Voice target and mode identified. Content loaded. Proceed only when gate passes.

Phase 2: SCAN

Goal: Run full checklist against content and identify all violations with evidence.

Step 0: Run deterministic stylometry checks

When the target voice has a profile.json, run the stylometry script first. It emits structured findings (rule_id, span, severity) that anchor the scan in measured data:

python3 scripts/voice-stylometry.py check \
  --profile skills/voice-{name}/profile.json \
  --draft <content-file>

Deterministic checks it runs:

• Burstiness band (burstiness.band, warning): draft sentence-length variance must fall inside the author's measured band; uniform sentence length is an AI tell.
• Punctuation profile (punctuation.em_dash|semicolon|parenthetical, warning): em-dash, semicolon, and parenthetical rates classified never/rare/habitual; flags drafts whose class deviates from the author's.
• Corrective antithesis (ai_tell.corrective_antithesis, error): "not X, it's Y" constructions, inline and across sentence pairs.
• Temporal openers (ai_tell.temporal_opener, error): throat-clearing paragraph openers ("In today's...", "In an era...", "Now more than ever...").
• Uniform paragraph shapes (ai_tell.uniform_paragraphs, error): four or more consecutive paragraphs with identical sentence counts.
• Profile decay (profile.stale, advisory): profile older than its refresh_after_days window. Advisory only — it asks for a profile refresh and never blocks the draft. Exit code stays 0 when only advisory findings exist.

Exit code 1 means error/warning findings exist; carry each finding into the violation list below. Profiles without stylometry or decay fields skip those checks and remain valid.

Step 1: Run negative prompt checklist

Check all categories against the target voice's checklist. Standard categories include:

• Tone: Does the tone match the voice profile? (e.g., too polished, too corporate, missing warmth)
• Structure: Does the structure match? (e.g., front-loaded constraints, clean outlines, wrap-ups)
• Sentences: Do sentence patterns match? (e.g., dramatic short sentences, rhetorical flourishes, symmetrical structure)
• Language: Any banned words? (amazing, terrible, revolutionary, perfect, game-changing, transformative, incredible, outstanding, exceptional, groundbreaking), marketing/hype, inspirational, unnecessary superlatives
• Emotion: Does emotion handling match? (e.g., explicitly named emotions, venting/ranting, moralizing)
• Questions: Do question patterns match? (e.g., open-ended brainstorming, vague curiosity)
• Metaphors: Do metaphor patterns match? (e.g., journey/path, biological/growth, narrative/story)
• Narrative (500+ words only; full 13-check rubric in references/narrative-patterns.md):
• Evidence speaks for itself, or narrator over-explains the point?
• Emotional registers mixed (named feelings + behavioral cues + sensation)?
• Secondary threads present that enrich the main argument?
• Reader's experience acknowledged?
• Temporal structure varied (callbacks, non-linear points)?
• References specific (named people, works, sources)?
• Room for uncertainty and loose ends, or subject has every answer?
• Intensity varies (peaks and valleys), or flat throughout?
• Event types diverse, or every paragraph does the same thing?
• Ending carries forward motion, or defaults to tidy realization epilogue?

Step 2: Check pass conditions

Verify the content matches the target voice's positive identity markers. Common pass conditions include:

• Feels like the person actually wrote it
• Voice-specific patterns are present (thinking out loud, warmth, precision, etc.)
• Could NOT be posted on LinkedIn without edits (for casual voices) — this heuristic catches ~80% of voice violations
• Does NOT sound like AI wrote it
• Narrative structure shows variety — secondary threads, temporal variation, forward-motion ending
• Mode-specific patterns are present (casual modes: no preamble, no wrap-up; formal modes: structured flow)

Step 3: Document violations

For each violation, record:

• Category (tone, structure, sentence, language, emotion, question, metaphor, narrative)
• Quoted text from the content
• Specific fix recommendation

Key constraint: Only scan at this stage; save revisions for the next phase. Subjective assessment without a checklist misses specific violations.

Gate: Full checklist scanned. All violations documented with evidence. Proceed only when gate passes.

Phase 3: REVISE

Goal: Fix all violations while preserving content intent and substance.

Step 1: Apply fixes

• Address each violation with the smallest change that resolves it
• Preserve the original meaning and information
• Maintain natural flow — fixes should not create new violations

Step 2: Verify no overcorrection

• Ensure revisions did not strip necessary content
• Confirm the substance and technical accuracy remain intact
• Keep the revision to voice-level changes only; leave paragraphs and arguments intact

Key constraint: Make the smallest change that resolves each violation. Preserve all meaning. Changing substance is scope creep.

Gate: All documented violations addressed. Intent preserved. Proceed only when gate passes.

Phase 4: VERIFY

Goal: Confirm revised content passes all checks.

Step 1: Rescan revised content

Run the full checklist from Phase 2 against the revised version.

Step 2: Evaluate result

• If PASS: Output final content with validation report
• If FAIL and iteration < 3: Return to Phase 3 with new violations
• If FAIL and iteration = 3: Output content with flagged remaining concerns

Key constraint: Always rescan. "Should be fine" is a rationalization. Fixes can introduce new violations.

Step 3: Output validation report

VOICE VALIDATION: [Voice Name] Mode [mode]
SCAN RESULT: [PASS/FAIL]
VIOLATIONS DETECTED: [N]
ITERATION: [1-3]

[If violations:]
1. [Category]: "[quoted violation]"
   Fix: [specific correction]

2. [Category]: "[quoted violation]"
   Fix: [specific correction]

REVISED OUTPUT:
[Corrected content]

RESCAN RESULT: [PASS/FAIL]

Gate: Content passes all checks, or maximum iterations reached with flagged concerns. Validation complete.

Examples

Example 1: Technical Voice Validation

User says: "Validate this draft is in the right voice"

Actions:

• Identify target voice from context, determine mode from content style (IDENTIFY TARGET)
• Run full 8-category checklist (Tone, Structure, Sentences, Language, Emotion, Questions, Metaphors, Narrative), find 2 violations (SCAN)
• Fix "I'm excited to share" (named emotion) and "This changes everything" (dramatic short sentence) (REVISE)
• Rescan revised content, confirm PASS (VERIFY)

Result: Clean content with validation report

Example 2: Community Voice Validation

User says: "Does this sound like the right voice?"

Actions:

• Identify target voice from context (IDENTIFY TARGET)
• Scan against voice checklist, find missing warmth and no sensory details (SCAN)
• Add experiential language and warmth while preserving substance (REVISE)
• Rescan, confirm warmth and sensory details present, PASS (VERIFY)

Result: Content matches voice profile

Error Handling

Error: "Voice Target Unclear"

Cause: Content doesn't specify which voice to validate against, or context is ambiguous

Solution:

• Check conversation context for voice mentions
• Look for voice-specific patterns to infer target
• If still unclear, ask user to specify voice name and mode

Error: "Violations Persist After 3 Iterations"

Cause: Fundamental mismatch between content substance and voice requirements, or conflicting checklist items

Solution:

• Output content with clearly flagged remaining violations
• List specific checklist items that resist correction
• Suggest the content may need to be regenerated from scratch with the correct voice skill

Error: "Revision Introduced New Violations"

Cause: Fixing one category created violations in another (e.g., removing dramatic sentences introduced polished phrasing)

Solution:

• Address new violations in next iteration
• If oscillating between two violation types, fix both simultaneously
• Prioritize tone and language violations over structural ones

References

Related Skills

• voice-{name} - Generates content in a specific voice (validate output with this skill)
• private de-AI editor skill - complementary anti-AI pattern detection
• voice-writer - Unified voice content generation pipeline that invokes this skill