What Song Refiners Actually Does Under the Hood
When we say Song Refiners uses AI, we're not talking about feeding your song to a generic chatbot and hoping for the best. We've built a specialized, purpose-built AI review system with nearly 6,000 lines of custom code -- plus over 100 backend functions -- that shape how the AI thinks, speaks, and evaluates your work.
Here's what's actually running when you hit "Submit."
Our Custom AI Framework
Genre-Specific A&R Personas for Every Genre
The AI doesn't approach every song the same way. We've programmed distinct "A&R personas" for every genre and sub-genre. Whatever genre you tell us, that's the lens the reviewer uses.
Each persona carries unique vocabulary, industry references, and evaluation criteria. A Christian ballad gets evaluated with encouragement and spiritual context. A Hip-Hop track gets judged on bars, flow, and quotables. An EDM submission focuses on the drop and festival-readiness. A Folk/Acoustic song is measured by storytelling craft and whether it would work with just voice and guitar.
The Conversational Intake
Before the AI ever listens to your song, our intake chatbot walks you through six focused questions: the backstory behind your song, your genre, how the lyrics were created, how the music was produced, what stage the recording is in, and your goals for the song. This conversation becomes the brief your reviewer works from -- so the feedback is grounded in YOUR context, not assumptions.
Tool-Based Creative Disclosure
Instead of asking you to guess "what percentage is AI-generated" on a slider, we ask you to select the specific tools you used -- Suno, ChatGPT, a co-writer, a DAW, live instruments. The system calculates authorship context from those selections, making the process honest and unambiguous. Selecting "Suno" is a clear signal; dragging a slider to 95% is not.
Explicit Instructions We Give the AI
Our system enforces dozens of explicit guidelines, including:
"Coach, Don't Prescribe" Philosophy (The Refiner's Fire)
The AI is NEVER allowed to write replacement lyrics, suggest specific titles, or prescribe melodic changes. Instead, it identifies WHAT isn't working, explains WHY, then asks coaching questions like "What image captures the feeling without dating the song?" to unlock YOUR creativity. Every feedback section ends with 2-3 mandatory coaching questions (4-5 for members with Double Fire reviews).
Tone Calibration
Every critique is framed as "untapped potential," not failure. The AI is prohibited from using condescending setups or belittling phrases. It's trained to root for you, not tear you down.
Scoring Calibration
We define exactly what each score range means -- and 100 doesn't exist. The ceiling is 99, because no song is perfect.
90-99: Exceptional -- would make industry professionals take notice
80-89: Very strong -- minor refinements needed
70-79: Solid but doesn't yet demand repeat listens
60-69: Promising foundation, needs development
Below 50: Needs fundamental rethinking
A score of 70 means "competent but unremarkable." To score above 75, the song must have something genuinely distinctive -- a fresh perspective, unexpected twist, or standout moment. We don't reward mere competence; we reward excellence.
Context-Aware Review Modes
The AI adapts its entire evaluation framework based on your submission:
Rough Demo Mode: Ignores recording quality entirely. Focuses on songwriting potential and gives a production roadmap for when you're ready to record or generate with Suno.
AI-Generated Music Mode: Evaluates creative direction and songwriting craft without penalizing AI vocal characteristics or synthetic sound qualities.
Studio Production Mode: Applies full professional production standards -- mix quality, mastering, performances.
Lyrics-Only Mode: Evaluates lyrics against the 10 Principles of Great Songwriting as if performed by a capable vocalist with the right melody.
Instrumental Mode: Evaluates sync licensing potential, melodic hooks, emotional arc, and composition craft for film, TV, and game placements.
Metaphor Intelligence
Songwriting is poetry set to music. When the AI encounters metaphors, allegory, or abstract imagery, its first job is to interpret what you're trying to express -- not flag it as "unclear." A line like "I'm drowning in a field of flowers" isn't contradictory; the tension is the point. The AI evaluates metaphors by whether they land emotionally, not whether they're literal.
Emotional Intent Recognition
If you share your song's backstory or Story Arc, the AI acknowledges the emotional weight while evaluating whether that emotion transmits to a listener who only has the lyrics and music -- not the backstory. Feedback is framed as "how to make the listener feel what YOU feel."
Banned AI-isms
We explicitly prohibit generic AI phrases:
"I'd be happy to help you..."
"Certainly," "Absolutely" as filler
"Furthermore," "Moreover" as transitions
Corporate buzzwords like "leverage" and "optimize"
Starting multiple sentences with "The"
The reviewer sounds like a human music industry veteran, not a chatbot.
Section-Specific Feedback
The AI references specific sections of your song (Verse 1, Chorus, Bridge) rather than timestamps, providing 4-6 section-specific callouts highlighting both strengths and areas for improvement. For Lyrics-Only reviews, these become line-by-line references pointing to exact moments in the text.
Revision Tracking
When you resubmit a revised version, the AI compares your new version against the original, explicitly noting what improved and what still needs work. Scores are compared side by side so you can see exactly how far you've come.
The Technology Stack
We use Google's Gemini 3 Pro -- a multimodal AI that can actually listen to your audio file, analyze production quality, and evaluate melody alongside lyrics. This isn't text processing; it's genuine audio analysis combined with lyrical craft evaluation.
Zero-Failure Architecture
Behind the scenes, we run a 3-tier provider fallback: Gemini 3 Pro as primary, Gemini 2.5 Pro as first fallback, and a secondary proxy as last resort. If one provider triggers a safety filter or goes down, the system automatically switches to the next -- with provider health tracking, cooldown management, and automatic retry scheduling. The result: your review always gets delivered.
Inline Quality Control
Before your review is delivered, it passes through an automated QC audit that checks for hallucinated lyrics (quotes that don't actually appear in your song), score inconsistencies, and formatting issues. If the AI fabricates a lyric quote, the system auto-corrects it before you ever see it.
Four Distinct Review Engines
Deep Dive Reviews -- Audio + lyrics evaluated together across 6 scoring categories (Overall, Title, Top Line, Lyrics, Arrangement, Production/Creative Direction). Members get Double Fire enhanced reviews with 50% more analysis depth, 4-5 coaching questions per section, and exclusive industry insights.
Quick Taste Preview -- A fast first impression that catches the biggest, most actionable insights. Uses an encouraging coaching tone to welcome new users. Available for free with your first credit.
Lyrics-Only Reviews -- For pre-production lyric development, applying the 10 Principles of Great Songwriting. Evaluates your words as if performed by a skilled vocalist, so you're never penalized for not having a recording yet.
Instrumental Reviews -- For tracks without vocals. Focused on sync licensing potential, melodic memorability, arrangement craft, and emotional arc. Feedback is framed from a music supervisor's perspective.
Beyond Reviews
The Generator -- Turn your reviewed lyrics into a full AI-generated song. The system sends your lyrics to Suno and delivers two track variations for you to choose from.
The Catalyst -- Take any reviewed song and remix it with a new style, tempo, or arrangement. Experiment until the production matches your vision.
Voice Coaching -- Talk directly with an AI coach about your review, powered by ElevenLabs conversational AI. Available 24/7, the coach has your full review context and follows the same Refiner's Fire philosophy -- it will challenge you, not hand you answers.
The Bottom Line
Generic AI would tell you your song is "nice" or "needs work." Song Refiners' custom system tells you exactly why your chorus hook isn't landing, asks you to consider what stakes you can raise in Verse 2, and guides you toward your next draft -- without ever putting words in your mouth.
We've spent years encoding the wisdom of industry veterans into a system backed by nearly 6,000 lines of custom review logic, over 100 backend functions, a zero-failure architecture, and inline quality control that catches mistakes before you see them. It's available 24/7, costs a fraction of professional feedback, and never gets tired, biased by personal taste, or afraid to tell you the truth.