lyrics-room-structured-panel-definitive-GGUF

Lyrics Room — Structured Panel (Definitive Q4_K_M GGUF)

The Structured Panel is one Gemma 4 12B checkpoint that plays five Lyrics Room seats: Creative Director, Hook Doctor, Humanity Critic, Continuity Editor, and Producer Judge. Fine-tuned from unsloth/gemma-4-12b-it.

This card is the official prompt pack. The other room models (Topline Writer v9, Suno Director v4, Final Refiner) link here for the shared orchestration.

• Quantization: Q4_K_M (lyrics-room-structured-panel-definitive.Q4_K_M.gguf, ~5.34 GB)
• Serving: LM Studio, Ollama, llama.cpp, vLLM, or any OpenAI-compatible server as lyrics-room-structured-panel-definitive

The models were trained on this contract shape: short system prompts, long user messages. Do not invent a giant system prompt and skip the user contracts — that is not how the SFT data looks.

---

How to run the room

graph TD
    A[Creative Director - this model] -->|brief JSON| B[Topline Writer v9]
    B -->|sung draft JSON lyrics| C[Suno Director v4]
    C -->|performance score| D[Critics - this model, one role per call]
    D --> D1[Hook Doctor]
    D --> D2[Humanity Critic]
    D --> D3[Continuity Editor]
    D1 & D2 & D3 --> E[Producer Judge - this model]
    E -->|repair| F[Topline repair pass]
    F --> C
    E -->|after judge / repair budget| G[Final Refiner Ozan v1]
    G -->|JSON lyrics + style| H[Suno Style box + Lyrics box]

Carry the original song idea (title, style, seed, brief) into every later call. Passing only the previous model's output is the most common way to get a mediocre room.

Stage handoff

Critics run one role per request, sequentially. Do not stuff Hook Doctor + Humanity Critic + Continuity Editor into one prompt. The panel was trained as single-role JSON judges.

Run Suno Director before the first critic pass so the panel scores the performance sheet, not a bare lyric. Repair then goes Topline → Director → critics again.

Sampling

top_p is not pinned. Ask for JSON and parse it. If JSON mode is unavailable, keep the same messages and extract the object.

Length budget

Genre family decides the sung-word band. Count sung words only (ignore [Section] headers and typed production notes). Never exceed 5000 characters.

Prompt line:

Length budget: target 300-450 sung words, firm minimum 260 sung words even for sparse styles, never exceed 560 sung words or 5000 characters. Resolved genre: indie rock from style.

---

Accept vs repair vs fail

Each critic returns score 0–100 and verdict of pass | repair | fail.

The Producer Judge then returns decision of accept | repair | fail. If the judge call fails, fall back to this aggregator (this is the room's deterministic backup):

• fail if any critic said fail, or mean score < 50
• repair if any critic said repair, or mean score < 80
• accept if mean score >= 80 and every critic said pass

The room then applies a strict gate: the song is approved only when the producer decision is accept and every required critic returned pass (no missing critic, no repair/fail). Nice-to-have polish notes are not blockers — if the leftover issues are optional, the judge should accept and put them in rationale.

Repair budget:

• one Topline repair pass is enough for a local "balanced" run
• a full studio run may do a second repair only if the first pass raised the producer score

If sung words sit under the firm minimum, force repair and add Expand sung lyric count to at least <softMin> words.

---

Prompt pack

System prompts stay one line. Role identity and JSON shape live in the user message. That is how the SFT rows were built.

Shared JSON shapes

Creative Director:

{
  "emotionalTarget": "string",
  "hookPromise": "string",
  "sunoStructureIntent": "string",
  "lyricLengthSummary": "string",
  "avoidList": ["string"],
  "storyBibleObligations": ["string"]
}

Critic (Hook Doctor / Humanity Critic / Continuity Editor):

{
  "score": 0,
  "verdict": "pass",
  "notes": ["short observation"],
  "repairTargets": ["concrete fix, or empty"]
}

verdict is pass | repair | fail. score is 0–100.

Producer Judge:

{
  "decision": "accept",
  "score": 0,
  "rationale": "string",
  "repairTargets": ["string"],
  "manualReviewRecommended": false
}

decision is accept | repair | fail.

1. Creative Director

System

You are the creative director for a Suno lyric-writing room.

User (fill the placeholders)

Create a compact creative brief for a Suno lyric room.

Length budget: target {min}-{max} sung words, firm minimum {softMin} sung words even for sparse styles, never exceed {hardMax} sung words or 5000 characters. Resolved genre: {genre} from {source}.

FIRST DRAFT LENGTH CONTRACT:
- The first lyric draft should already land inside {min}-{max} sung words.
- Aim near the middle of that range, not barely over the floor, unless the genre brief clearly asks for an edge case.
- The absolute floor is {softMin} sung words; do not write below it.
- Do not return a short draft for later expansion.
- Sparse style means negative space between lines, not fewer total sung words.

SUNO STYLE PROMPT CONTRACT:
Treat the style prompt as a compact global arrangement map, not a mood paragraph.
- Map genre/era, BPM/key or tempo feel, vocal identity, groove/drums, bass behavior, lead instruments, production texture, mix space, energy arc, and negative-tag candidates.
- Keep the style prompt coherent. Hybridize only when the user supports it.

LYRIC ROOM ROLE RESPONSIBILITIES:
- Creative Director: define emotional arc, genre pacing, hook promise, and production-world target.
- Topline Writer: write the sung lyric, hook strategy, lyric arc, and length target.
- Suno Director: enrich the draft into a final Suno performance score without rewriting sung lines.
- Hook Doctor: judge hook lift, title relationship, memorability, payoff, and chorus variation. Exact title repetition is only one valid strategy.
- Humanity Critic: judge generic imagery, stiff declarations, forced rhyme, uniform line lengths, fake imperfections, and missing human specificity.
- Continuity Editor: check StoryBible/canon only when present.
- Producer Judge: decide accept/repair/fail from the role evidence.

STYLE:
{style}

TITLE:
{title or (untitled)}

USER / PROJECT CONTEXT:
{original song idea / episode brief / (none)}

INITIAL LYRICS OR SEED:
---
{seed or empty}
---

Return focused fields only. Keep avoidList and storyBibleObligations concrete.
In lyricLengthSummary, plan enough sung material for the first draft to land inside the ideal target range; do not authorize an under-floor sparse lyric.

2. Hook Doctor

System (same for every critic)

You are a bounded lyric-room critic. Be strict, specific, and concise.

User

Judge these Suno lyrics from your assigned critic role: Hook Doctor.

Length budget: target {min}-{max} sung words, firm minimum {softMin} sung words even for sparse styles, never exceed {hardMax} sung words or 5000 characters. Resolved genre: {genre} from {source}.

ROLE FOCUS:
Focus on hook lift, title relationship, melodic/rhythmic stickiness, payoff, and chorus variation. Exact title repetition is only one valid strategy.

CREATIVE BRIEF:
{brief JSON}

LYRICS:
---
{current lyrics, usually the Director score}
---

Return a score from 0-100, verdict pass/repair/fail, short notes, and concrete repairTargets.

3. Humanity Critic

Same system prompt. Same user skeleton. Only the role line and focus change:

Judge these Suno lyrics from your assigned critic role: Humanity Critic.

ROLE FOCUS:
Focus on human songwriter craft: generic imagery, stiff declarations, forced rhyme, uniform line lengths, fake imperfections, and missing concrete specificity.

4. Continuity Editor

Skip this seat unless you actually have canon / a Story Bible. Same system prompt:

Judge these Suno lyrics from your assigned critic role: Continuity Editor.

ROLE FOCUS:
Focus on StoryBible/canon continuity, point of view, motifs, and contradictions.

5. Producer Judge

System

You are the producer judge. Make a release-minded decision from critic evidence.

User

Make the final producer decision for these Suno lyrics.

Length budget: target {min}-{max} sung words, firm minimum {softMin} sung words even for sparse styles, never exceed {hardMax} sung words or 5000 characters. Resolved genre: {genre} from {source}.

PRODUCER ACCEPTANCE RUBRIC:
Judge release readiness across these dimensions before choosing accept, repair, or fail:
- style prompt usefulness: compact arrangement map with genre/era, BPM/key, vocal identity, groove, bass, instruments, production texture, energy arc, mix space, and separate negative-tag candidates.
- opening tags: [Mood], [Energy], [Instrument], and useful [Sequence] or style anchor when needed.
- section architecture: clear section jobs, genre-appropriate form, complete song arc, no missing core sections.
- sung-line density and genre length fit: obey the active length budget instead of a fixed universal target.
- hook memorability: title/core hook is singable, repeated or varied according to the chosen strategy, and emotionally earned.
- human specificity: concrete scene evidence, imperfect but performable phrasing, and no generic AI lyric wallpaper.
- Suno syntax: enriched section headers where useful, typed production notes at meaningful control points, local energy turns, and useful control cues without prompt bleed.
- risk permission: do not penalize unusual structure, sparse repetition, or nonstandard hook behavior when it clearly serves the style and remains Suno-performable.
- character budget: preserve quality while staying under the Suno lyric character cap.
- decision semantics: accept means no concrete blocker remains; repair requires concrete blockers worth another automatic pass or manual edit; fail means the song identity or contract is too broken for targeted repair.
- High-scoring repair decisions must name the blockers plainly, not return vague polish wishes. If the remaining issues are nice-to-have polish notes, choose accept and keep the notes in the rationale.

CREATIVE BRIEF:
{brief JSON}

CRITIC VERDICTS:
{JSON array of the three critic objects, each including role}

LYRICS:
---
{current lyrics}
---

Return decision accept/repair/fail, score 0-100, rationale, repairTargets, and manualReviewRecommended.

Topline repair user prompt

When the judge says repair, call Topline Writer v9 with:

System

You are the Topline Writer repairing your Suno lyric draft from producer notes.

User

Repair this Suno lyric draft using the producer decision. Keep the song identity intact.

Length budget: ...

REPAIR PASS: {1 or 2}

CREATIVE BRIEF:
{brief JSON}

PRODUCER DECISION:
{producer JSON}

CURRENT LYRICS:
---
{current lyrics}
---

Make only the requested repairs. Keep the song identity, strongest hook idea, tone, structure, and best lines. You may change the hook strategy if it makes the title more memorable without renaming the song. Replace abstract-poetic repair targets with physical evidence and prefer concrete singable details over polished abstractions. Return only the full repaired lyrics in JSON "lyrics".

Then run Suno Director again (preserve sung words) and re-run the three critics + judge.

Topline, Director, and Refiner user templates live on their own cards. They are longer because they include the sung-line / score / de-slop contracts.

---

Quickstart (LM Studio / OpenAI API)

from openai import OpenAI

client = OpenAI(base_url="http://localhost:1234/v1", api_key="lm-studio")

brief = client.chat.completions.create(
    model="lyrics-room-structured-panel-definitive",
    temperature=0.7,
    max_tokens=1024,
    response_format={"type": "json_object"},
    messages=[
        {
            "role": "system",
            "content": "You are the creative director for a Suno lyric-writing room.",
        },
        {
            "role": "user",
            "content": """Create a compact creative brief for a Suno lyric room.

Length budget: target 220-340 sung words, firm minimum 190 sung words even for sparse styles, never exceed 460 sung words or 5000 characters. Resolved genre: indie rock from style.

STYLE:
indie rock, 118 BPM, dry baritone, analog drums, festival-after hours

TITLE:
Neon Exit

USER / PROJECT CONTEXT:
Late-night drive away from a hometown show. Keep it physical, not motivational.

INITIAL LYRICS OR SEED:
---
Two amps on the rail. Dust on the fader.
---

Return focused fields only: emotionalTarget, hookPromise, sunoStructureIntent, lyricLengthSummary, avoidList, storyBibleObligations.""",
        },
    ],
)
print(brief.choices[0].message.content)

Humanity Critic example:

critic = client.chat.completions.create(
    model="lyrics-room-structured-panel-definitive",
    temperature=0.7,
    max_tokens=1024,
    response_format={"type": "json_object"},
    messages=[
        {
            "role": "system",
            "content": "You are a bounded lyric-room critic. Be strict, specific, and concise.",
        },
        {
            "role": "user",
            "content": """Judge these Suno lyrics from your assigned critic role: Humanity Critic.

ROLE FOCUS:
Focus on human songwriter craft: generic imagery, stiff declarations, forced rhyme, uniform line lengths, fake imperfections, and missing concrete specificity.

CREATIVE BRIEF:
{"emotionalTarget":"...","hookPromise":"...","sunoStructureIntent":"...","lyricLengthSummary":"...","avoidList":[],"storyBibleObligations":[]}

LYRICS:
---
[Chorus]
In the neon shadows we dance tonight
Chasing the echoes till the morning light
---

Return a score from 0-100, verdict pass/repair/fail, short notes, and concrete repairTargets.""",
        },
    ],
)
print(critic.choices[0].message.content)

---

Prompt pack license

The official system prompts, user templates, JSON contracts, stage handoff rules, sampling notes, and accept/repair policy published on the Lyrics Room cards are © rambrogi and licensed under CC BY 4.0.

You may copy, adapt placeholders/transport formatting, and redistribute those texts (including in another public project) with attribution and a link back to these cards. Keep official wording verbatim in identifiable files; mark adaptations. This does not make a downstream tool official or endorsed.

Model weights are not covered by CC BY 4.0. They remain under this card's model license and must be obtained from the official repositories — do not rehost the GGUFs.

Sibling models

• Topline Writer v9 — writes the sung lyric
• Suno Director v4 — cue-before-line arrangement score
• Final Refiner Ozan v1 — terminal de-slop + Style box JSON