audio-jingle

Files (2)

• SKILL.md
• example.html

SKILL.md (first lines)

# Audio Jingle Skill

Three sub-modes. The active project's `audioKind` decides which one
runs:

| `audioKind` | Models we route to | Plan focus |
|---|---|---|
| `music` | Suno V5 (default), Udio, Lyria 2 | genre + tempo + instrumentation |
| `speech` | MiniMax TTS (default), Fish, ElevenLabs V3 | script + voice + pacing |
| `sfx` | ElevenLabs SFX (default), AudioCraft | texture + impact + duration |

## Resource map

```
audio-jingle/
├── SKILL.md
└── example.html
```

## Workflow

### Step 0 — Read the project metadata

`audioKind`, `audioModel`, `audioDuration` (seconds), and (for speech)
`voice`. Branch by known values and use them verbatim. Missing metadata is not
an instruction to ask: infer a safe default when possible, and emit a
clarifying form only when the missing answer would materially change the
requested output or prevent generation.

Important: `voice` is provider-specific. For `minimax-tts`, `--voice`
must be a valid MiniMax `voice_id` (for example `male-qn-qingse`), not
a natural-language description. If you only have a prose voice brief
("warm female narrator", "neutral Mandarin"), keep that in your plan
but omit `--voice` so the daemon's default voice id applies, or ask the
user to choose a specific id.

### Step 1 — Plan

**Music**
- Genre + reference artists (1-2)
- Tempo (BPM) + key
- Instrumentation (3-5 instruments max)
- Vocals: yes / no / hummed / choir
- Mood arc (intro → chorus → outro)

**Speech**
- Script (final, not draft — TTS runs verbatim)
- Voice target + pacing
  For MiniMax this means a real `voice_id`, not prose in `--voice`
- Pronunciation hints for proper nouns / acronyms

**SFX**
- Texture (impact / whoosh / ambience / foley)
- Duration + envelope (sharp attack vs. gentle swell)
- Layering note (single hit vs. stacked)

State the plan in 2-3 sentences before dispatching.

### Step 2 — Compose the prompt

Use the format the upstream model prefers. Bind `audioDuration` to the
API parameter directly; never put "make it 30 seconds" in prose.

### Step 3 — Dispatch via the media contract

Use the unified dispatcher — do **not** call provider APIs by hand:

```bash
"$OD_NODE_BIN" "$OD_BIN" media generate \
  --project "$OD_PROJECT_ID" \
  --surface audio \
  --audio-kind "<music|speech|sfx>" \
  --model "<audioModel from metadata>" \
  --duration <audioDuration seconds> \
  [--voice "<provider voice id (speech only)>"] \
  --output "<short-slug>-<duration>s.mp3" \
  --prompt "<assembled prompt from Step 2 — for speech, the literal script>"
```

The command prints one line of JSON: `{"file": {"name": "...", ...}}`.
The bytes land in the project; the FileViewer renders the audio
transport controls automatically.

### Step 4 — Hand off

Reply with: plan summary, the filename returned by the dispatcher, and
one sentence on what to try if the user wants a variation (e.g. "swap
tempo from 92 to 108 BPM" rather than "make it different").

## Hard rules

- TTS runs your script **literally**. Proof it before dispatching —
  even one stray comma changes the cadence.
- MiniMax TTS rejects free-form voice prose in `--voice`. Use a real
  MiniMax `voice_id` (for example `male-qn-qingse`) or omit the flag
  and let the daemon's default voice apply.
- Music: under 30s = single section; 30–90s = intro + body; 90s+ =
  full arc. Don't try to fit a 3-act song into 15 seconds.
- SFX: prefer one well-described layer over a paragraph of "make it
  cool" — generators reward specific texture words.
- Save the file every turn. The audio viewer shows transport controls
  the moment the file lands.

From hiktan44/open-design-app, Apache-2.0.