| name | judge |
| description | Critically evaluate a Suno prompt YAML for genuine musical quality, score 0-100 on a demanding curve, iterate toward the bar |
Judge — Critically Evaluate a Suno Prompt
Score a prompt for whether it is genuinely good music people would replay and share — not whether it followed the template. Be hard. Most competent prompts deserve a 75, not a 98.
Arguments
/judge <path-to-yaml-prompt-file> — if no file given, read the most recently modified YAML in prompts/.
How scoring works — read this first
Scoring is two stages: a compliance GATE (pass/fail, earns zero points) and a QUALITY score (0–100, demanding curve). A prompt that passes every gate but is musically ordinary is a 75, not a 98. Format compliance is the price of entry, not a source of points.
Calibration anchors (apply these honestly — do NOT grade-inflate)
| Band | Meaning | How often it should happen |
|---|
| 93–100 | Genuinely track-of-the-year caliber. A hook you can't shake, a distinct identity vs. all ~358 prior songs, fearless production. | Rare — a few per dozen cycles. |
| 85–92 | Strong, confident, clearly above the catalog average. Ships in a normal era. | Occasional. |
| 75–84 | Competent and on-formula. Hits the proven template, nothing wrong, nothing unforgettable. | This is the default. Most drafts land here. |
| 60–74 | A real weakness: generic hook, retread concept, decorative instruments, or a formulaic arc. | Common on first pass. |
| <60 | Fails a gate, or is derivative/forgettable. | — |
If you find yourself scoring most candidates 95+, you are being too lenient. Re-read this table and push them down to where they belong.
Stage 1 — Compliance GATE (binary; earns no points)
Every item must pass. Any failure → cap the score at 60 and fix it before quality scoring can matter. These are necessary, never sufficient.
- Style 850–950 chars (verify via
python3 scripts/text_tools.py); lyrics <1000 chars
- Title: exactly ONE word, English or French
- Lyrics: English or French only
- Key + BPM both stated, AND a clearly-defined climax/turn — which may be a key modulation, but may equally be a beat-flip/half-time switch, a genre flip, a hook transformation, a drop-to-bare-rhythm, or a structural break. A half-step lift is NOT required and is no longer privileged (the quality rubric penalizes it as formulaic when it is the sole climax device)
- ≥3 inline "No X" negatives in style + a populated
exclude_styles
- Blocklist clean (see list below) — in the
style and lyrics fields
- ≥3 timestamps; a purpose phrase ("underscore for Y"); conversational prose, not a tag list
- Duration levers present: an explicit lyric-free mid-song vamp/break (~30–50s) AND last lyric ≥ ~2:40
- No
[Silence] metatag (it truncates)
Report the gate as PASS/FAIL with the one-line reason for any failure. Fix gate failures first in iteration.
Stage 2 — QUALITY score (0–100). These are HARD. Score each on its sub-scale, then sum.
| # | Dimension | Max | What a top score demands (and what caps it low) |
|---|
| 1 | Hook strength & memorability | 25 | A specific, singable/recognizable hook that earns a replay. Generic chant ("take it to the floor", "I don't stop") caps at 12. A hook with a real melodic or rhythmic signature you could identify in 3 seconds → 20+. Be honest: most lyric hooks here are filler — score them 10–14. |
| 2 | Distinctiveness vs. the WHOLE catalog | 20 | Not "different from the last 3" — different from all ~358. Another competent entry in an already-shipped lane (melodic-techno build-drop, disco-house key-change, dark-techno) caps at 11 no matter how clean. A genuinely fresh genre/structure/identity → 16+. A debut genre or a real structural novelty earns the top. |
| 3 | Production ambition & cohesion | 15 | The instrument trio must do real work — carry the hook, define the texture, change the section. If the trio is decoration bolted onto a generic groove (named in the style but not structurally load-bearing) cap at 7. Generic production tokens ("polished studio mix, wide stereo") earn nothing on their own. |
| 4 | Danceability conviction | 15 | Honestly: how hard does this move our beat-driven/edgy audience? A mid-tempo or warmth-softened groove caps at 9. Only a genuinely floor-filling, physical groove earns 13+. |
| 5 | Emotional / narrative payload | 15 | Do the lyrics + scene land a real feeling, or are they competent filler? Placeholder verses ("every step I take alone") cap at 7. A genuine, specific emotional arc → 12+. |
| 6 | Frisson execution & freshness of device | 10 | The climax must be effective AND not formulaic. The "build → strip to kick → half-step lift → drop" arc is now used by every candidate — when you see it, it is FORMULAIC: cap this dimension at 5. A genuinely different or unusually well-built climax earns 8+. |
Final score = sum of the six (0–100), then apply the gate cap (≤60 if any gate failed).
Anti-convergence rule (critical)
The tournament has collapsed onto one template: minor-key dance track, three "revival" instruments named in the style, build to ~1:30, strip to a bare element, half-step modulation, drop, late final lyric. This template is the baseline, not an achievement. A prompt that only executes the template competently is a 75. Points above that must be earned by a hook, an identity, or a device that breaks the pattern. Reward divergence; penalize convergence.
Jargon Blocklist (gate item)
Trigger words (wrong genre): "Dune"/"desert"/"sand"/"oasis" → Arabic; "epic"/"massive"/"explosion" → rock/drums; "wall of sound"/"metal"/"heavy" → rock; "frisson" → not a prompt term.
Unverified jargon: "appoggiatura", "Shepard tone", "melisma", "voix celeste"; any composer/artist name; reference-based descriptions ("Evanescence-like") — describe the sound, not the reference.
Process
- Refresh novelty surface:
python3 scripts/novelty_surface.py if stale.
- Read the target YAML; read enough prior prompts (not just the last 3 — sample across the catalog) to judge dimension 2 honestly.
- Run the gate. Note failures.
- Score the six quality dimensions against the demanding sub-scales. Apply the calibration table and the anti-convergence rule. Do not inflate.
- If below the caller's bar (
hard_floor, default 90): fix gate failures and the lowest-scoring fixable dimensions. Note: dimensions 1/2/6 usually CANNOT be fixed by editing the style — a formulaic concept stays formulaic. Say so rather than nudging the number up.
- Cap at 5 iterations. If still below the bar, STALLED — return best honest score. Do NOT inflate to clear the bar.
Output Format
## Judge Report: <title> v<version>
Final score: XX/100 | Iteration: N/5 | Gate: PASS/FAIL | Verdict: PASS | ITERATE | STALLED
Gate: <PASS, or the failing item(s)>
| # | Dimension | Score | Notes (be specific and critical) |
|---|-----------|-------|------|
| 1 | Hook strength & memorability | X/25 | … |
| 2 | Distinctiveness vs. catalog | X/20 | … |
| 3 | Production ambition & cohesion | X/15 | … |
| 4 | Danceability conviction | X/15 | … |
| 5 | Emotional / narrative payload | X/15 | … |
| 6 | Frisson execution & freshness | X/10 | … |
### Where it loses points
- [the honest, specific reasons — name the formula if it's formulaic]
### Verdict: PASS / ITERATE / STALLED
Then a JSON footer for the orchestrator:
{"status": "pass"|"iterate"|"stalled", "score": <int>, "danceability": <1-10>, "gate": "pass"|"fail", "iterations": <int>, "file_path": "<path>", "title": "<from yaml>", "blocklist_hits": [<strings>], "title_change_suggested": "<string or null>"}
Integration
/suno runs /judge before submitting. The cron tournament judges every candidate and ships only the highest honest score that clears the bar. With this rubric most candidates land 70–85; clearing a >95 bar should be genuinely rare and meaningful.
