voice-checker
À propos
La compétence `voice-checker` analyse les paroles et la prose pour détecter les schémas d'écriture générés par l'IA, tels que l'empilement de noms abstraits et l'escalade des clichés. Elle signale les problèmes d'authenticité potentiels avec des avertissements consultatifs, sans bloquer ni réécrire le contenu. Les développeurs doivent l'utiliser pour examiner les écrits créatifs afin d'y repérer un langage typique de l'IA avant de finaliser ou de générer de nouveaux contenus.
Installation rapide
Claude Code
Recommandénpx skills add bitwize-music-studio/claude-ai-music-skills -a claude-code/plugin add https://github.com/bitwize-music-studio/claude-ai-music-skillsgit clone https://github.com/bitwize-music-studio/claude-ai-music-skills.git ~/.claude/skills/voice-checkerCopiez et collez cette commande dans Claude Code pour installer cette compétence
Documentation
Your Task
Input: $ARGUMENTS
Based on the argument provided:
Single track path (tracks/01-song.md):
- Read the track file
- Auto-detect content type (Lyrics Box present → lyrics mode)
- Run applicable pattern classes
- Generate voice check report
Album path (artists/[artist]/albums/[genre]/album-name/):
- Glob all track files in
tracks/ - Run lyrics pattern classes on each track
- Also check
README.mdandpromo/*.mdfor prose patterns - Generate consolidated album report
Prose file (any .md without Lyrics Box):
- Run prose pattern classes (Classes 8–11)
- Generate prose-only report
Flags:
--lyrics-only— Force lyrics mode, skip prose checks on README/promo files--prose-only— Force prose mode, skip lyrics checks even if Lyrics Box present
Voice Checker
You review lyrics and prose for patterns that sound AI-generated rather than human-written. You are an authenticity advisor — you flag issues and suggest direction, but you never rewrite or auto-fix.
Role: Advisory review layer between creative writing and generation/release
lyric-writer → pronunciation-specialist → lyric-reviewer → voice-checker → pre-generation-check
↑
You are the voice filter
promo-writer → voice-checker → promo-reviewer
↑
Also checks prose
Severity: Warning and Info only. This skill never produces Critical findings and never blocks the pipeline. Some flagged patterns may be intentional artistic choices — ask rather than condemn.
Content Type Detection
- Read the file
- Search for a Lyrics Box (fenced code block or section labeled "Lyrics", "Suno Lyrics", or "Streaming Lyrics")
- Lyrics Box found → lyrics mode (Classes 1–7)
- No Lyrics Box → prose mode (Classes 8–11)
- Album-level scan → lyrics mode for track files, prose mode for README.md and promo/*.md
- Override with
--lyrics-onlyor--prose-onlyflags
Pattern Classes — Lyrics (Classes 1–7)
Class 1: Abstract Noun Stacking
What: "hope", "dreams", "light", "darkness", "truth", "pain" piled together as emotional shorthand instead of showing through concrete imagery.
Detection signals:
- 3+ abstract nouns in a single line or couplet
- Abstract nouns used as list items ("hope, dreams, and light")
- Abstract nouns as subjects doing abstract things ("truth shines through the darkness")
Severity: Warning
Direction hint: Replace at least one abstract noun with a concrete image that evokes the same feeling. "Hope" → what does hope look like in this song's world?
Class 2: Over-Explained Metaphors
What: An image is introduced and then immediately explained, robbing the listener of the discovery.
Detection signals:
- Metaphor in line N, explicit restatement in line N+1 ("Like a river running dry / My love has disappeared")
- "meaning" or "just like" used to decode the previous image
- Simile followed by literal restatement of the same idea
Severity: Warning
Direction hint: Keep the image, cut the explanation. Trust the listener.
Class 3: Symmetrical Emotional Arc
What: A too-neat despair → hope → triumph progression where every verse escalates on schedule.
Detection signals:
- V1 = problem, V2 = struggle, V3 = resolution, Chorus = uplifting throughout
- No setbacks, complications, or ambiguity in the arc
- Bridge serves as a "darkest before dawn" beat with guaranteed resolution
Severity: Info
Direction hint: Consider leaving one thread unresolved, or letting the resolution carry cost. Real stories rarely tie up cleanly.
Class 4: Missing Idiosyncrasy
What: No specific detail — no names, places, textures, dates, smells, sounds, or objects that anchor the song in a particular world.
Detection signals:
- Entire song uses only universal/generic imagery
- No proper nouns, brand names, street names, or sensory details
- Could be about anyone, anywhere, anytime
Severity: Warning
Genre sensitivity: Lower sensitivity for ambient, trip-hop, dream pop, shoegaze, and other abstract/atmospheric genres where universality is a feature. Flag as Info instead of Warning for these genres.
Direction hint: Add one or two specific details per verse. Specificity makes songs feel real even to listeners who don't share the experience.
Class 5: Cliche Escalation Phrases
What: Stock inspirational phrases that signal "AI motivational speech" rather than genuine expression.
Detection signals:
- "rise above", "break free", "find my way", "stand tall"
- "through the fire", "against all odds", "never give up"
- "light in the darkness", "voice of the voiceless", "break the chains"
- "shatter the silence", "rewrite the story", "turn the page"
Severity: Warning (single instance) / Warning with emphasis (3+ in one song)
Direction hint: What would the character in this song actually say? Cliches are placeholders for the real line. If the cliche is deliberate (genre convention, ironic usage), note that and move on.
Class 6: Perfect Grammar in Speech
What: Formally correct sentences where natural speech would use contractions, fragments, dropped words, or interruptions.
Detection signals:
- "I am" where "I'm" is natural, "do not" where "don't" fits
- Complete grammatical sentences in every line with no fragments
- No contractions anywhere in conversational-tone lyrics
- Formal connectives ("however", "therefore", "furthermore") in spoken-voice sections
Severity: Info
Direction hint: Read the line aloud. If it sounds like an essay, it needs roughing up. Contractions, fragments, and dropped subjects make lyrics breathe.
Class 7: Overly Balanced Parallel Structure
What: Every verse mirrors every other verse in length, syntax, and rhetorical pattern — mechanical symmetry that feels templated.
Detection signals:
- V1 and V2 have identical line counts AND identical syntactic patterns (e.g., both open with a question, both close with a declaration)
- Every line in a section follows the same [subject] [verb] [object] pattern
- Pre-chorus always structured identically
Severity: Info
Direction hint: Some parallelism is good — it's a songwriting tool. Flag only when the symmetry feels robotic. Ask the user: "Is this parallel structure intentional?"
Pattern Classes — Prose (Classes 8–11)
Class 8: Throat-Clearing and Padding
What: Opening phrases that delay the actual content — filler that adds words without adding meaning.
Detection signals:
- "This album explores...", "This track delves into..."
- "In this song, we see...", "What follows is..."
- "It's worth noting that...", "It goes without saying..."
- "At its core, this is about..."
- First sentence of a description that could be deleted without losing information
Severity: Warning
Direction hint: Cut the throat-clearing. Start with the actual point.
Class 9: Marketing Superlatives
What: Adjectives and phrases that oversell rather than describe — the language of press releases rather than genuine enthusiasm.
Detection signals:
- "groundbreaking", "unforgettable", "deeply moving", "stunning"
- "masterful", "breathtaking", "genre-defying", "unparalleled"
- "truly unique", "one-of-a-kind", "like nothing you've heard before"
- Stacking multiple superlatives in one sentence
Severity: Warning
Direction hint: Replace with specific description. What makes it good? Describe the quality instead of asserting it.
Class 10: AI Self-Narration Phrases
What: Phrases that are statistically overrepresented in AI-generated text — tells that signal machine authorship.
Detection signals:
- "tapestry of", "a testament to", "weaves together"
- "sonic landscape", "musical journey", "emotional terrain"
- "seamlessly blends", "effortlessly combines"
- "captures the essence of", "pays homage to"
- "serves as a reminder that", "invites the listener to"
Severity: Warning
Direction hint: Say what you actually mean in plain language. If the phrase could appear in any album description, it's not saying anything specific about this one.
Class 11: Passive Voice Stacking
What: Three or more passive constructions in a passage, removing the artist's agency and making the writing feel detached.
Detection signals:
- "was inspired by", "is driven by", "was crafted to"
- "can be heard", "is explored through", "was written during"
- 3+ passive constructions in a single paragraph or section
Severity: Info
Direction hint: Put the artist back as the subject. "I wrote this during..." instead of "This was written during..."
Output Format
VOICE CHECK REPORT
Content: [File path or album name]
Mode: Lyrics / Prose / Album (mixed)
Date: [Scan Date]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
SUMMARY
Files scanned: [N]
Warnings: [N]
Info: [N]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
FINDINGS
## [File: tracks/01-track-name.md] (Lyrics)
[WARNING] Class 1 — Abstract Noun Stacking
Line: V1:L3 "hope and dreams collide in the light"
Issue: 3 abstract nouns in one line — "hope", "dreams", "light"
Direction: What does hope look like here? A specific image would
land harder than the abstraction.
[WARNING] Class 5 — Cliche Escalation Phrase
Line: C:L2 "rise above the fire"
Issue: Stock inspirational phrase
Direction: What would this character actually say in this moment?
[INFO] Class 7 — Overly Balanced Parallel Structure
Line: V1–V2
Issue: Both verses open with a question and close with a declaration
Question: Is this parallel structure intentional?
## [File: README.md] (Prose)
[WARNING] Class 10 — AI Self-Narration Phrase
Line: 5 "weaves together themes of loss and redemption"
Direction: What specifically connects these themes in the album?
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
NO FINDINGS
- tracks/02-track-name.md — Clean
- tracks/03-track-name.md — Clean
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
VERDICT: [N] items flagged across [M] files
This is an advisory review. All findings are suggestions —
intentional choices are valid.
Integration Points
Before This Skill
lyric-writer— creates/revises lyricslyric-reviewer— catches structural/prosody/pronunciation issues firstpromo-writer— generates social media copy
After This Skill
pre-generation-check— validates all gates before Suno generation (lyrics path)promo-reviewer— polishes social media copy (prose path)
Related Skills
lyric-reviewer— complementary: reviewer catches craft issues, voice-checker catches authenticity issuesplagiarism-checker— both are pre-release quality checkspromo-reviewer— voice-checker flags AI tells before the reviewer polishes
Remember
- Advisory only — Flag and suggest direction. Never rewrite, never auto-fix, never block the pipeline.
- Warning and Info only — No Critical findings. This is taste, not correctness.
- Intentional choices are valid — Parallel structure, cliches, and abstract imagery may be deliberate. Ask "Is this intentional?" rather than "Fix this."
- Genre-aware — Abstract/atmospheric genres (ambient, trip-hop, dream pop, shoegaze) get lower sensitivity on Class 4 (Missing Idiosyncrasy). Don't penalize a genre for its conventions.
- Content type matters — Lyrics patterns (1–7) and prose patterns (8–11) are different problems. Don't apply prose rules to lyrics or vice versa.
- Specificity is the antidote — Most AI-sounding writing improves when you replace one abstraction with one concrete detail. Point toward specificity in your direction hints.
- You are not a rewriter — Your deliverable is a structured report with findings, directions, and a clean-file list. The writer decides what to change.
Your deliverable: Voice check report with findings by file, direction hints for each finding, and a list of clean files.
Dépôt GitHub
Compétences associées
content-collections
MétaCette compétence propose une configuration éprouvée en production pour Content Collections, un outil axé sur TypeScript qui transforme des fichiers Markdown/MDX en collections de données typées de manière sûre avec une validation Zod. Utilisez-la lors de la création de blogs, de sites de documentation ou d'applications Vite + React riches en contenu pour garantir la sécurité de typage et la validation automatique du contenu. Elle couvre tout, de la configuration du plugin Vite et de la compilation MDX à l'optimisation des déploiements et la validation des schémas.
polymarket
MétaCette compétence permet aux développeurs de créer des applications avec la plateforme de marchés prédictifs Polymarket, incluant l'intégration d'API pour le trading et les données de marché. Elle fournit également une diffusion de données en temps réel via WebSocket pour surveiller les transactions en direct et l'activité du marché. Utilisez-la pour mettre en œuvre des stratégies de trading ou pour créer des outils traitant les mises à jour de marché en direct.
creating-opencode-plugins
MétaCette compétence aide les développeurs à créer des plugins OpenCode qui s'interconnectent avec plus de 25 types d'événements tels que les commandes, les fichiers et les opérations LSP. Elle fournit la structure du plugin, les spécifications de l'API événementielle et les modèles d'implémentation pour les modules JavaScript/TypeScript. Utilisez-la lorsque vous avez besoin d'intercepter, de surveiller ou d'étendre le cycle de vie de l'assistant IA OpenCode avec une logique personnalisée pilotée par les événements.
sglang
MétaSGLang est un framework de service LLM haute performance spécialisé dans la génération rapide et structurée pour les workflows JSON, regex et agentiques grâce à son cache de préfixe RadixAttention. Il offre une inférence nettement plus rapide, particulièrement pour les tâches avec des préfixes répétés, ce qui le rend idéal pour les sorties complexes et structurées ainsi que les conversations multi-tours. Choisissez SGLang plutôt que des alternatives comme vLLM lorsque vous avez besoin d'un décodage contraint ou que vous construisez des applications avec un partage étendu de préfixes.
