Buddy Sings

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 & 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:

1. 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.

2. 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.

3. 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