Vibe Sing

When the user invokes /vibe-sing, run the pipeline script. It distills the current session's emotional vibe into a Lyria music prompt (via Gemini) and composes a song.

What to do

1. Run the skill's launcher, passing through the user's argument verbatim:

``bash ~/.claude/skills/vibe-sing/run.sh ``

Recognized arguments:

• (none) — 30-second clip (default)
• pro — ~2 minute full song
• stop — kill the currently-playing song

1. The script handles everything: finding the current session's transcript JSONL, slicing the recent vibe-relevant text (skipping tool calls/results), calling Gemini for a mood prompt, calling Lyria, saving the mp3 to ~/.claude/skills/vibe-sing/generations/, and open-ing the file so it plays.

1. After the script returns, the last line of stdout is a JSON object with mood_prompt, audio_file, model, and gemini_model. Report back to the user in at most two sentences:

• One sentence quoting the mood prompt Gemini chose.
• One sentence with the file path.

Do not describe what the song is "about" — let the audio speak. Do not add commentary about whether the vibe seems right. The user will hear it and decide.

Errors

If the script fails (missing GOOGLE_API_KEY, no transcript, API error), surface the stderr message verbatim and stop — do not retry.

Notes

• Requires GOOGLE_API_KEY in .env inside the skill dir (or in the environment).
• The transcript reader explicitly skips tool calls and tool results to keep the vibe extraction focused on conversational tone.
• Both the transcript filter and the Gemini system prompt are designed to prevent corny "song about debugging React" output — the mood prompt should never reference programming, files, or specific topics.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: harajlim
• Source: harajlim/vibe-sing
• License: MIT

Install and usage instructions live in the source repository linked above.