SKILL.md

$27

Audio player (for playback — at least one of):

mpv (recommended — interactive controls: space = pause, q = quit)

ffplay (from FFmpeg)

afplay (macOS built-in)

Workflow Overview

Check pet → Build vocal identity → Gather context → Generate music → Play &#x26; feedback

Language & Interaction

Detect the user's language from their first message. Respond in the same language

throughout the entire session. All examples below are in English — translate them

naturally when responding in other languages.

User-facing text localization rule:

ALL text shown to the user — including pet info, voice description, lyrics preview,

prompt preview, playback info, and feedback prompts — MUST be fully translated into

the user's language.

The API prompt sent to the model should always be written in English for best

generation quality. However, when previewing the prompt to the user, show a localized

description in the user's language instead of the raw English prompt. The English prompt

is an internal implementation detail — the user does not need to see it.

The templates below are written in English as reference. At runtime, translate every

label and message into the user's detected language.

The pet sings in the user's language by default. Embed the singing language

naturally in the vocal description (e.g., "singing in Japanese" or "singing in

Mandarin Chinese") rather than appending a separate language tag. If the user

explicitly requests a different language for the lyrics, honor that request.

Step 1: Check for Pet

Read ~/.claude.json and look for the companion field.

If no companion is found or the field is empty, tell the user:

You don't have a pet yet! Type /buddy to adopt one, then come back to let it sing.

Stop here and wait for the user to adopt a pet. Do not proceed without a pet.

If a companion exists, extract its profile:

name — the pet's name

personality — the pet's personality description

Present the pet to the user:

Found your pet!

   Name: <name>

   Personality: <personality>

Step 2: Build Vocal Identity

Based on the pet's name and personality text, creatively design a unique

vocal identity. No template lookups — interpret the personality freely.

How to interpret personality into voice

Read the personality text and craft vocal attributes:

Timbre: What does this personality sound like? e.g., "few words" →

low, warm, deliberate; "energetic" → bright, punchy; "mysterious" → breathy,

dark; "legendary chonk" → thick, warm, cozy

Singing style: How would they deliver a song? e.g., "of few words" →

sparse, dramatic pauses; "playful" → bouncy, rhythmic; "poetic" → flowing, legato

Mood: What emotional tone fits? e.g., "chill" → relaxed, laid-back;

"fierce" → intense, powerful

Construct a prompt_fragment that describes the vocal style in English, embedding

the singing language naturally. For example:

Vocal: warm low female voice singing in Mandarin Chinese with cozy thick timbre,

sparse minimalist delivery with dramatic pauses giving each word weight, relaxed

laid-back mood.

Voice caching

The vocal identity must be cached so the pet always sounds the same.

Cache file: ~/.claude/skills/buddy-sings/voices/<name>.json

Cache format:

{

  "name": "Moth",

  "personality": "A legendary chonk of few words.",

  "prompt_fragment": "Vocal: warm low female voice singing in Mandarin Chinese...",

  "cached_at": "2026-04-07T19:52:15"

}

First time: No cache exists → interpret personality → save to cache file.

Subsequent calls: Read cache → use the saved prompt_fragment directly.

Do NOT re-interpret — consistency matters.

Cache invalidation: If the personality in ~/.claude.json differs from what's

cached, the pet has changed — regenerate and save a new cache.

Manual regeneration: If the user says "change the voice" or "regenerate voice":

delete the cache file and re-interpret from scratch.

Present the voice to the user

<name>'s unique voice:

Timbre: <timbre description>

Style: <style description>

Mood: <mood description>

Let's pick what <name> should sing about!

Step 3: Understand Intent & Gather Context

Do NOT always present a mode menu. Instead, analyze the user's request to

determine what context is needed, and auto-gather it.

Auto-context detection

When the user's request implies personal context, automatically scan for

relevant information without asking. Triggers include:

Time-based references: "today", "this week", "recently", "yesterday" → scan

current conversation history and memory files for what happened in that period

Personal references: "my work", "my day", "what I did" → scan memory

and conversation for the user's activities

Relationship references: "our story", "what we did together" → scan memory for

shared experiences between user and pet/Claude

Context gathering (auto, not mode-gated)

When context is needed, scan these sources in order:

Current conversation context: Look at what the user has been doing in

this Claude Code session — files edited, commands run, topics discussed.

This is the richest source for "today" type requests.

Memory files: Scan for relevant memories:

find ~/.claude/projects/*/memory/ -name "*.md" 2>/dev/null | head -20

Also check ~/.claude/memory/ if it exists.

Read found files and extract themes relevant to the user's request.

Git history (if in a repo): For work-related songs, check recent commits:

git log --oneline --since="today" 2>/dev/null | head -10

Use gathered context to enrich the lyrics prompt — make the song personal and

specific to what actually happened, not generic.

When NO context is needed

If the user's request is a clear standalone scene (e.g., "sing a rainy day song",

"sing a lullaby"), skip context gathering and proceed directly to music generation.

When context is ambiguous

Only ask for clarification when you genuinely can't determine what the user wants.

Don't present a mode menu — ask a specific question:

What should <name> sing about?

For example:

  - "Today's work" — I'll check what you've been up to

  - "My pet waiting by the window for me to come home"

  - Or let me pick a random theme?

Fallback to random

If context gathering finds nothing useful (no memory files, no conversation

history, no git log), fall back to random theme generation based on the pet's

personality:

Quiet/reserved personality → midnight lullaby, gentle sunset, quiet morning

Energetic personality → party jam, adventure song, victory march

Mysterious personality → moonlit serenade, secret whisper, dream journey

Tell the user what theme was picked.

Step 4: Generate Music

Combine the vocal identity with the chosen theme.

Construct the full prompt: The prompt has two parts that MUST both be present:

Part A — Vocal identity (MUST come first): Always start the prompt with the

cached prompt_fragment. This is the most important part — it defines who is

singing. Place it at the beginning of the prompt so the API prioritizes it.

Part B — Genre/style/mood tags: Choose tags that match the theme, NOT

a default set. Vary the genre deliberately based on what the song is about.

Write prompts as vivid English sentences, not comma-separated tags.

Follow this pattern: A [mood] [genre] song, featuring [vocal description], about [narrative/theme], [atmosphere], [key instruments and production].

Describe vocals as a character ("sultry baritone with jazz inflections"),

not just a gender. Include a scene or vibe to anchor the generation.

Genre matching guidelines — pick a genre that fits the theme's energy:

Theme energy

Suggested genres

Avoid

Encouragement / motivation / cheer

Indie rock, synth-pop, funk, rap

Indie folk, healing

Daily life / warmth / companionship

Mandopop, city pop, bossa nova

Same as last time

Missing someone / waiting

Folk, R&B, lo-fi

Rock, EDM

Humor / roasting / complaining

Funk, rap, ska, electro-pop

Classical, ballad

Late night / quiet

Ambient, piano piece, lo-fi, neoclassical

Upbeat, EDM

Celebration / achievement

EDM, future bass, funk, K-pop

Slow tempo, melancholy

Work routine

City pop, synth-pop, lo-fi hip-hop, indie rock

Same genre every time

Anti-monotony rule: NEVER use the same genre combination twice in a row.

Before constructing the prompt, recall what genre was used in the previous

generation (if any in this session) and pick something different.

Prompt structure — write as vivid English sentences, not comma-separated tags:

<vocal prompt_fragment>. A <genre> song with <mood> mood, featuring <instruments>,

at a <tempo> tempo, evoking <scene>.

Diverse examples:

# Encouragement for the workday

A deep warm androgynous voice with cozy delivery. An energetic synth-pop track

with a fiery, uplifting mood, driven by pulsing synthesizers and electronic drums

at a fast tempo, capturing the rush of a morning commute.

# Waiting for the owner to come home

A deep warm androgynous voice with cozy delivery. A warm city pop song with sweet,

tender feelings, featuring electric piano and groovy bass at a mid-tempo pace,

set on a sunny afternoon windowsill waiting for someone to come home.

# Complaining about overtime

A deep warm androgynous voice with cozy delivery. A playful funk track with a

humorous, laid-back vibe, featuring slap bass and brass at a groovy mid-tempo,

capturing the absurdity of working late in a dim office.

# Late-night companionship

A deep warm androgynous voice with cozy delivery. A calm lo-fi hip-hop piece with

a healing, dreamy atmosphere, featuring sampled piano and soft electronic drums

at a slow tempo, evoking a quiet late-night desk with warm lamp light.

Generate lyrics: Use --lyrics-optimizer to auto-generate lyrics, or write lyrics

yourself when you need to control the perspective.

Important — perspective & personality-driven lyrics:

The pet is the singer, so lyrics MUST be written from the **pet's first-person

perspective** ("I" = the pet, "you" = the owner/user). The pet is singing TO

the owner. For example:

"I sit by the door waiting for you to come home" (pet's perspective)

"Wake up now, my dear human" (pet singing to owner)

NOT "I rub my sleepy eyes" (owner's perspective — wrong)

NOT "Then you woke up, my little Moth" (owner talking about pet — wrong)

The pet's personality should shape the lyrics' tone and word choice:

"of few words" → short, impactful lines, minimal filler

"playful" → rhyming, bouncy phrasing, fun wordplay

"poetic" → metaphor-rich, flowing imagery

"fierce" → direct, powerful declarations

The pet's name may appear in the lyrics (e.g., in a chorus hook) but the

narrative voice is always the pet speaking/singing.

When perspective matters: Write the lyrics yourself and pass via --lyrics.

When perspective is not critical: Use --lyrics-optimizer for convenience.

Preview (MUST show full content): Before generating, show the user the

complete lyrics and full prompt — no abbreviation, no ..., no summary.

This is part of the fun — the user wants to read and enjoy the lyrics before

hearing them sung.

The API prompt is always constructed in English (for best generation quality).

When responding in a non-English language, show a localized description of

the prompt in the user's language for readability. The English prompt is an

internal implementation detail — do not show it to the user. Translate all

labels (Singer, Theme, Description, Confirm, etc.) into the user's language.

Template (English reference — localize all labels at runtime):

About to generate:

Singer: <name>

Theme: <theme>

Lyrics:

[verse]

<full verse lyrics here>

[chorus]

<full chorus lyrics here>

... (show ALL sections in full)

Description: <localized description of the song style and mood>

Confirm? (press enter to confirm, or tell me what to change)

Never truncate or abbreviate the lyrics or prompt in the preview.

The user should see exactly what will be sent to the API.

Call music generation:

With auto-generated lyrics (perspective not critical):

mmx music generate \

  --prompt "<full combined prompt>" \

  --lyrics-optimizer \

  --out ~/Music/minimax-gen/<name>_sings_<YYYYMMDD_HHMMSS>.mp3 \

  --quiet --non-interactive

With self-written lyrics (perspective-controlled):

mmx music generate \

  --prompt "<full combined prompt>" \

  --lyrics "<lyrics with correct pet perspective>" \

  --out ~/Music/minimax-gen/<name>_sings_<YYYYMMDD_HHMMSS>.mp3 \

  --quiet --non-interactive

Step 5: Play & Feedback

Cross-platform playback

Detect the available audio player and play the generated file:

if command -v mpv >/dev/null 2>&#x26;1; then

  mpv --no-video ~/Music/minimax-gen/<filename>.mp3

elif command -v ffplay >/dev/null 2>&#x26;1; then

  ffplay -nodisp -autoexit ~/Music/minimax-gen/<filename>.mp3

elif command -v afplay >/dev/null 2>&#x26;1; then

  afplay ~/Music/minimax-gen/<filename>.mp3

else

  echo "No audio player found. Your song is saved at: ~/Music/minimax-gen/<filename>.mp3"

fi

After starting playback, tell the user the file is playing and where it's saved.

Do NOT show playback controls (e.g. keyboard shortcuts) — they don't work in this

environment since the player runs in the background.

If no player is found, show the file path and suggest installing mpv.

Feedback

After playback, ask for feedback (localize all text):

How was <name>'s performance?

1. Amazing! Keep it!

2. Try a different theme / style

3. Fine-tune the lyrics and regenerate

4. Try another random one

Edge Cases

Situation

Action

No ~/.claude.json

Tell user to run /buddy first

Companion field is empty

Same — guide to /buddy

mmx CLI not installed

Print: "You need to install mmx CLI: npm install -g mmx-cli && mmx auth login"

No audio player found

Show file path and suggest installing mpv

No memory files found

Suggest a custom theme or random mode

User wants to change the pet's voice

Delete cache, re-interpret personality

User wants a specific genre

Let them override — append their genre to the prompt

Notes

The vocal identity is based on name + personality only. No species/rarity

template mapping.

Voice is cached and consistent across sessions. Same pet = same voice.

Lyrics should always be original — never reproduce copyrighted lyrics.

The pet's personality shapes both the voice (how they sound) and the

lyrics (what they say and how they say it).

All generated files go to ~/Music/minimax-gen/ with the pet name in the

filename.

buddy-sings