Agent Skills Specification

The audio-stemming skill complies with the open Agent Skills Specification.

Three-Stage Progressive Disclosure

1. Discovery Stage: At agent startup, only the YAML frontmatter name and description are ingested into context. Because the description is constrained to 278 characters, agents retain this capability with negligible token overhead.
2. Activation Stage: When a prompt matches the skill’s capabilities (e.g., “isolate vocals from track.wav”), the agent loads the full instructions from SKILL.md.
3. Execution Stage: The agent follows the structured instructions, referencing the decision matrix and invoking bundled runner scripts.

Discovery & Installation Paths

To support both single-plugin and monorepo host discovery:

• Antigravity & Claude Code: Looks for skills/<skill-name>/SKILL.md directly at the repository root. This is supported via the top-level skills symlink:

  skills/audio-stemming/SKILL.md -> plugins/audio-stemming/skills/audio-stemming/SKILL.md

• Global Agent Discovery: Link the skill directly into your host environment:

  ln -s "$(pwd)/skills/audio-stemming" ~/.gemini/config/skills/audio-stemming
  ln -s "$(pwd)/skills/audio-stemming" ~/.claude/skills/audio-stemming

Skill Package Structure

plugins/audio-stemming/skills/audio-stemming/
├── SKILL.md                          # Frontmatter + operational instructions
├── scripts/
│   ├── separate.sh                   # Non-interactive CLI wrapper
│   └── benchmark.py                  # Host qualification checker
├── references/
│   ├── stemmer-matrix.md             # Model trade-offs and memory ceilings
│   └── error-catalog.md              # Actionable remediation codes
└── assets/
    └── sample_meta.json              # Sample telemetry output schema

Frontmatter Metadata

---
name: audio-stemming
description: Separate mixed audio tracks into isolated vocal and instrumental stems (vocals, drums, bass, guitar, piano, other) using local HTDemucs PyTorch inference. Use when an agent needs to isolate vocals, strip drums or accompaniment, or extract multitrack stems from WAV or MP3 files.
license: Apache-2.0
compatibility: Requires uv, Python >= 3.11, ffmpeg, and >= 3GB available RAM.
metadata:
  author: ghchinoy
  version: "0.1.0"
  tags: ["audio", "stems", "htdemucs", "music-processing", "separation"]
---