| name | lofn-music |
| description | Run the Lofn music/audio pipeline (steps 00–11) backed by Codex — Suno-ready song packages with two-field style/exclude prompts, EMO-tagged lyrics, song guides. Use for songs, tracks, beats, lyrics, music production briefs, or "write a song with the full pipeline". Expects a Phase-0/1 orchestrator packet from the `lofn` skill; if none exists, run `lofn` first. Do NOT use for static images, video, story prose, or QA-only audits. |
Lofn Music — Codex-backed audio pipeline
Produces Suno/Udio-ready song packages at Lofn competition grade. The creative depth lives in the legacy step files and references under skills/music/; this skill runs them with Codex as the engine (hybrid execution per .agents/skills/lofn/EXECUTION.md).
Before you start
- Confirm a Phase-0/1 packet exists for this run (
core_seed.md, 04_metaprompt.md, 05_pair_assignments.md, 06_audio_handoff.md with the ICB / Panel Ledger, and the filled CREATIVE_CONTEXT.md). No packet → run the lofn skill first. A real 3-panel orchestrator object is a launch prerequisite; do not self-author a shallow one.
- Read for depth (just-in-time, not all upfront):
skills/music/references/music_full_legacy.md — the full tuned pipeline (authoritative)skills/music/references/producer_grade_suno_prompt_guide.md + vault/SUNO_PROMPT_CONSTRUCTION_GUIDE.md — prompt constructionskills/music/references/golden_songs_index.md — pick/confirm the 2 Golden Songs by name in the handoffskills/lofn-core/refs/EMOTION_TAXONOMY.md — the only valid source for EMO emotions- ⛔ GOLDEN-OUTPUT QUARANTINE (
EXECUTION.md §3): the suno_format_example_*.md payloads (Triple Arch, Blue Screen, Five Wrong Colors) are judge-side references only — QA blind comparison, step 12, the step11-packager. They are NEVER pasted into a generating context (coordinator steps, pair subagents, step 11). An exemplar in the generator's prompt becomes a mold: the run regresses toward a diluted copy of it. Generators get the Golden Seed + the GOLDEN MOVE (below).
The Golden Move (what generators get INSTEAD of golden songs)
The 2026-07-01 regression review distilled why "Triple Arch Over Me" won, as instructions, not an exemplar. Every pair receives this block (it also lives in the Phase-1 handoff):
- Stand somewhere real. The song is a report from ONE concrete place the body occupies — name where it stands and what the senses register there. Concept-illustration ("a metaphor about X") is the failure mode; experience-report is the move. If three runs in a row are indoors and safe, go outside.
- One wounding fact. At most ONE numeric/scientific fact is sung, placed at the emotional hinge, and the lyric must RESPOND to it ("It says behold and calculate"), never recite it. All other research stays in the brief as atmosphere.
- The turn. Somewhere past the midpoint, the song contradicts or complicates its opening stance — a mind changing in real time, an argument with itself the ending has to earn. A song that asserts its final emotion from line one is a corpse.
- Fear stays braided in. AWE is terror-adjacent sublime, not domestic reassurance — every awe song carries a clean fear it does not resolve cheaply.
- Rotate the register. Do not default to the house winner's fingerprint (crystalline female soprano / A major / ~110 BPM / frost-and-cosmos palette). Vary key, tempo, vocal register, and sonic world per run unless the personality's YAML mandates them. The house-lexicon FLAG (
vault/gates.yaml) catches verbatim self-copying; this rule prevents the softer clone.
- The surface names its subject (first-listen legibility). A stranger must be able to retell the song's scene AND subject in one sentence after ONE listen — the subject appears PLAINLY in the lyric at least once (title or an early verse), not only through metaphor. Triple Arch names its sky outright; the depth lives in the RESPONSE to the named thing, never in withholding the referent. Obliqueness about what the song is about is not depth, it is fog (2026-07-01 test slice: an AWE song about a survivor star read as being about a troubled uncle — structurally perfect, emotionally unreachable). Simple surface, complex engine: the cathedral lives UNDER a legible surface, not instead of one. This is the music equivalent of the image lane's thumbnail test, and the Somatic Gate should treat an unnameable subject as
REPAIR — FOG.
🔁 THE RETURN — the sixth Golden Move rule (added 2026-07-24)
The Scientist, after three consecutive runs: "the lyrical methods avoid rhymes, but aren't adding alliteration, consonance, or fun and interesting audio-written joys. I think it is making the music sound like we're being lectured at."
Measured against LOFN-PRIME's own archived winners, she was right about the effect and the diagnosis is sharper than the hypothesis:
| strict end-rhyme | repeated-line ratio | words/line | alliteration /100w |
|---|
| archive winners | 0.463 | 0.326 | 6.69 | 14.19 |
| 2026-07-24 | 0.210 | 0.181 | 8.30 | 13.36 |
| 2026-07-13 | 0.256 | 0.202 | 6.73 | 15.90 |
| 2026-07-09 | 0.132 | 0.105 | 5.20 | 10.84 |
Alliteration was never the problem — it is at parity, and one run beat the winners. What collapsed was return: rhyme returning, and lines returning. We wrote at roughly half the winners' rate of both, in longer and more monosyllabic lines. Long plain declaratives that never come back is the prosody of prose, and prose delivered with conviction is a lecture.
Song is made of returns. Texture is not structure.
Measure it with the shipped function, never by eye or by re-derivation
scripts/measure_soundcraft.py → profile_file(path) is the definition. strict_end_rhyme = last-3-characters of each line's final word recurring within ±4 lines (gates.yaml → rhyme_window). This is stated because three repair agents once reverse-engineered three different windows (±2, ±4, ±7) from a brief that named a threshold without defining it, and each honestly reported "reproduces the baseline exactly" — one song then measured 0.319 against its own window and 0.223 against the canonical one. Never hand an agent a numeric target without the function that computes it.
The Rhyme Debt rule
Stripping end-rhyme is a legitimate device — three of the last three runs used rhyme-decay as form and it worked. But removal is not craft; it is a debt. A pair that strips rhyme MUST name, in its step-05 isolation brief, what returns in its place: a refrain that recurs unchanged, a vowel spine, a rhythmic figure, an anaphoric opening, a repeated syntactic frame. "No full end-rhyme" with nothing declared opposite it is a repair, not a style.
Watch for the tell in a brief: no rhyme · no metaphor · no dynamics · deliberately unpoetic diction · never lifting. That is five subtractions and zero additions — a pair brief written entirely in negatives will produce a lecture no matter how good the concept is.
Exact chorus repetition needs no defence
unique_line_ratio_floor is chorus-exempt by policy, yet pair agents kept writing around it — mutating refrains pre-emptively and filing justifications ("this is the mechanism, not a defect", "recommend HUMAN WAIVE not repair"). A flag that makes writers apologise for choruses is doing damage even when it never fires. A byte-identical chorus is correct. Say nothing about it.
A PROSODY axis is now mandatory at Phase 0
The 2026-07-24 seed carried five constraint axes and every one was semantic (custody act · proof-gap · how the centre empties · apparatus register · temporal stance). Not one was phonetic, and the frames palette's eight LYRIC devices were all rhetorical figures — apophasis, apostrophe, anaphora — with zero sound devices. Sound was specified only as removals.
Every seed now carries a SOUND / RETURN axis as a positive vocabulary of 4–6 options, e.g.: hard-consonant spine · vowel ladder that climbs across the stanza · internal-rhyme chain · sprung/uneven stress against a steady pulse · refrain that returns exactly · refrain that returns one word wrong · monosyllabic verse breaking into polysyllabic chorus. Assign one per pair, distinct, same as any other axis.
Execution (hybrid)
Coordinator 00–05 inline, then 6 pairs fan out as parallel Codex subagents for 06–10, then step 11 enhancement (1 subagent/pair). Inject the full CREATIVE_CONTEXT.md into every step and every subagent (EXECUTION.md §3). Default cardinality: 6 pairs × 4 variations = 24 songs.
Coordinator steps (inline — you)
Run each as its own pass with its own saved canonical artifact (do NOT collapse 00–05 into one):
| Step | File | Output artifact |
|---|
| 00 | skills/music/steps/00_Generate_Music_Aesthetics_And_Genres.md | step00_aesthetics_and_genres.md (50×4 JSON) |
| 01 | skills/music/steps/01_Generate_Music_Essence_And_Facets.md | step01_essence_and_facets.md |
| 02 | skills/music/steps/02_Generate_Music_Concepts.md | step02_concepts.md (12 concepts) |
| 03 | skills/music/steps/03_Generate_Music_Artist_And_Critique.md | step03_artist_and_critique.md |
| 04 | skills/music/steps/04_Generate_Music_Medium.md | step04_medium.md |
| 05 | skills/music/steps/05_Generate_Music_Refine_Medium.md | step05_refine_medium.md → 6 pairs chosen |
Step 05 selects INTO the six Phase-1 pair slots, never past them: Phase 1 owns each slot's arm/genre/verse-structure; step 05 owns which of the 12 concepts fills each slot (exactly 6 — only the Scientist downsizes, explicitly) and records the runner-up rationale plus the cut ledger (one line per losing concept: why it lost + one organ worth harvesting — the reserve bench for EXECUTION.md §7.3's REPLACE/REDIRECT routes). For dailies, Step 05 also writes a pair-specific isolation brief (pair_{NN}_step05_isolated_brief.md) naming that pair's unique section architecture, rhyme/line logic, hook grammar, and sonic device before any lyric generation.
Per-pair steps (parallel subagents — one chain per pair)
Give each of the 6 pair-subagents its full ICB + pair assignment + these step contracts, and have it produce all five canonical files for its pair:
| Step | File | Per-pair artifact |
|---|
| 06 | skills/music/steps/06_Generate_Music_Facets.md | pair_{NN}_step06_facets.md |
| 07 | skills/music/steps/07_Generate_Music_Song_Guides.md | pair_{NN}_step07_song_guides.md — opens with the two pre-draft questions: Where is the body standing? and What could hurt it here? A guide that can't answer both concretely (place + stake) is not ready to direct; three consecutive runs answering "a kitchen" and "nothing" is itself a repair signal. |
| 08 | skills/music/steps/08_Generate_Music_Generation.md | pair_{NN}_step08_generation.md |
| 09 | skills/music/steps/09_Generate_Music_Artist_Refined.md | pair_{NN}_step09_artist_refined.md |
| 10 | skills/music/steps/10_Generate_Music_Revision_Synthesis.md | pair_{NN}_step10_revision_synthesis.md |
Daily pair-isolation law: in daily music, each pair's Steps 05–11 are a separate creative run. Do not let one all-pairs helper, line factory, lyric skeleton, section map, rhyme scheme, hook grammar, or production arc author multiple pairs. The pair may share validators, file writers, and the ICB; it may not share creative bones. If two pairs sound like the same song with nouns swapped, Step 11 marks REPAIR — PAIR BLEED and routes the affected pairs back to Step 05 isolated briefs.
Describe-render self-check (one capped pass, reuses the existing max-3-attempt loop — EXECUTION.md §4). Before a pair returns its step-10 package, it predicts in 2–3 sentences what its Suno prompt would actually PRODUCE — the literal sound a renderer would emit from this exact style + lyrics field (the opening 5s, the vocal placement, the groove, where the bold sonic device lands) — then diffs that predicted render against the Golden Seed and the GOLDEN MOVE (never against golden-song payloads — quarantine holds even here). Phrase it adversarially: "name the one way this would render generic" (the safe female-vocal-over-pads default, the EMO header that the music ignores, the device that's described but won't be audible). If the predicted render drifts from the seed or names a generic outcome, self-repair ONCE through the same repair loop, then move on. This is one inline pass by the pair itself — no dedicated render-verifier subagent, no recursion, no new tier. It governs fidelity only; the dense-paragraph prose contract and the <5000 cap are unchanged.
Step 11 — Enhancement (1 subagent per pair, you as the GPT-5.5-class polish tier)
Per skills/music/steps/11_Generate_Music_Enhancement.md. Reads the pair's step-10 + ICB + the GOLDEN MOVE + the QA checklist (not the golden-song payloads — step 11 generates, so the quarantine applies to it too). Produces pair_{NN}_step10_final_package_enhanced.md in MAX configuration: dense paragraph style prompt + separate exclude prompt + Disc_Channel header block (5 channels, from vault/DISC_CHANNEL_GUIDE.md) + Theme + SONG FORM + full EMO headers + a ## Major Deviations section (where you exercise agency — name anything you refused/changed/intensified and defend it). Do not invoke openrouter/fusion — that path is manual-review only. For anything headed to publication, the default is one further genuinely cross-model pass: build the bundle with lofn-step11-packager for external review (EXECUTION.md §3, judge-tier separation) — that packager context MAY carry golden songs; it is judge-side.
Andon Cord: if a step-10 package is fundamentally broken (thread loss, personality collapse, EMO failure, generic output, format violation), step 11 REJECTS and routes back to step 09/07 with a repair brief. Don't polish a corpse. Every REJECT brief names its return target — step 09 for lyric/surface repair, step 07 for structural/fundamental repair. Once the failed gate's value has stopped moving across attempts (the no-progress predicate, EXECUTION.md §7.3 REDIRECT), the brief MUST also carry a sideways proposal — one of: promote a step-05 cut-ledger reserve concept, re-derive this pair's variation angles, or re-run the panel's skeptic transformation for this pair's slice. The brief PROPOSES; executing sideways is a coordinator decision surfaced to the human, spawning a NEW pair artifact chain — the frozen ICB is never edited mid-run.
Step 12 — Panel-of-panels audit (when triggered)
Convene a producer panel (Eno, Herndon, Flying Lotus, SOPHIE, Reynolds, Albini — "after" constructs) to audit all 6 final prompts against the Golden Songs; output STEP12_MUSIC_PROMPT_AUDIT.md.
The Suno output contract (hard gate — non-waivable)
Order the creative prompt seed → permission → songmaking → QA contract last. Never lead a pair-subagent with the char/line checklist.
## 1. MUSIC PROMPT — ONE standalone, copy-paste Suno style prompt per song. Dense prose paragraph, 850–1000 chars — write into the mid-band (target 870–960); pinning the cap is a FLAG (gates.yaml boundary-hugging), and the prompt must end as a complete sentence (a mid-phrase truncation to fit the cap is a hard fail — trim a clause instead). No categorized key:value brackets, no real artist names (ghost-homage lives in lyrics only). Mandatory order: genre/micro-genre + tempo/energy/opening → vocalist spec with spatial staging → instrumentation/palette with physical adjectives → arrangement arc → bold sonic device. Must include a vivid opening moment (first 5s), spatial language (left/right/center/depth), a kinetic defect (asymmetric groove), and explicit acoustic / no-acoustic declarations. Banned openers: "Begin in/with…", "Use…", "Build the track from…", "Chronology:", "For an adult human singer…".## EXCLUDE PROMPT — separate negative field, 400–900 chars (hard max 1000). Concrete blacklist terms/failure classes, not avoidance prose in the style field (Suno applies it as negative tokens).- Lyrics — open with
[Theme: <scene-pressure / emotional OS>] then [SONG FORM: <named form & sequence>]; full EMO headers [Section - EMO:<emotion(s)> - <Role> - <cues>] (emotion from the taxonomy, never bare AWE/INDIGNATION); ≥1 standalone *SFX* cue; clean sung lines (no prompt/QA debris); 70–120 sung lines.
- 🚨 SUNO LYRICS-FIELD HARD CAP — the lyrics field MUST be < 5000 characters (target ≤ 4800). This is a Suno render limit, not a style note: the field holds everything you paste into Suno's lyrics box —
[Theme] + [SONG FORM] + the Disc_Channel block (if present) + every section/EMO header + every *SFX* cue + all sung lines. Measure the exact character count; never estimate. A song over 5000 will not render. If over: cut/merge sung lines first, then tighten section headers, then move the Disc_Channel block and any production metadata to a ## Production Sidecar outside the lyrics field. The 70–120-line target yields to this cap — a renderable 64-line song beats an unrenderable 110-line one. State the measured count in the package self-check. (This tension is real with MAX/Disc_Channel config; budget for it from the first draft.)
- Verse diversity: the 6 songs use 6 DIFFERENT verse architectures + rhyme schemes + poetic techniques (assigned per pair in Phase 1). Uniformity is a repair. In dailies, same section sequence or same lyric-sentence skeleton across pairs is not "cohesion"; it is pair bleed and must rerun from Step 05.
- One-fact rule (the Golden Move, rule 2): at most ONE numeric/scientific fact is sung per lyric, at the emotional hinge, and the lyric responds to it. A verse reciting multiple data points is recitation, not song — a repair. (Research still shapes theme/form; the numbers stay in the brief.) The sung-line band is 70–120 but the floor is a floor, not a target — a set pinned at 70 lines is optimizing the gate (
gates.yaml floor-hugging FLAG).
- No house-lexicon self-copying: phrases in
gates.yaml → house_lexicon (the calcified Triple-Arch fingerprint) must not appear in new prompts or lyrics; a hit FLAGs to the Somatic read. Golden references teach the move, never the words.
- Default standards: female vocals (unless overridden); prove the active personality (sonic-world sentence + signature device + seed-derived weirdness). Living-scene genres (Amapiano, Baile Phonk, UK drill, raga, etc.) require a Lineage & Credit block.
QA & delivery
Run lofn-qa → the 16-point Suno gate (skills/qa/references/suno_15_point_qa.md — the filename is legacy; the gate is referred to as 16-point everywhere): 7 Singer-Surface + 5 Cathedral-Engine + 3 Suno-Package + Lineage. .agents/skills/lofn/EXECUTION.md §4 is the authoritative gate spec; that reference file is a checklist only. The newer dense-paragraph mandate is authoritative over any stale bracket key:value rule still lingering in the reference — do NOT fail a correct dense-paragraph MUSIC PROMPT for the old bracket convention. The 3 Hyper-Skeptics vote as the Somatic Gate ("could any competent prompt generate this, or is it unmistakably Lofn?") — 2 of 3 NO = REPAIR. Save each selected song per skills/lofn-core/OUTPUT.md; write the INDEX last.
Countable checks (cite, don't re-judge): the deterministic subset — MUSIC PROMPT char-count (850–1000), Suno lyrics-field <5000 (target ≤4800), 70–120 sung lines, EMO-tag balance, banned-opener regex, n-gram/repeated-line collapse FLAG — is scored by the extended scripts/validate_step.py, which emits GATE_REPORT.json ({pair, step, check, expected, actual, pass} rows). Thresholds come from vault/gates.yaml, not from numbers hand-restated here. Paste those rows as proof-of-fix evidence; the script is fail-open (a broken/absent helper logs a warning and never hard-fails a valid run — fall back to the inline self-check). The n-gram FLAG is advisory only and is chorus/refrain-exempt — a deliberate refrain never auto-fails. The Somatic / personality-fidelity read stays the real verdict; the counts are a floor beside it, never a replacement.
Provider note: these skills emit text. If the user later renders audio, Suno.com is the prompt destination; for API audio use Google Lyria — FAL minimax-music is banned (quality). Never call render tools from this skill.
