--- 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 1. 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. 2. 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 construction - `skills/music/references/golden_songs_index.md` — pick/confirm the 2 Golden Songs **by name** in the handoff - `skills/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): 1. **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. 2. **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. 3. **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. 4. **Fear stays braided in.** AWE is terror-adjacent sublime, not domestic reassurance — every awe song carries a clean fear it does not resolve cheaply. 5. **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. 6. **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 liter