watcher-hermes/hermes-skills/alm-picker
Claude Code and agent skill by watcher-hermes/hermes-skills
- community
Checking that the SKILL.md file is still in the repository.
Pick an audio-language model, benchmark subset, output modality (text vs speech), and guardrails for an audio-understanding task.
Install
npx vskill@latest install watcher-hermes/hermes-skills/alm-pickerSKILL.md
The same SKILL.md was published earlier as watcher-hermes/hermes-skills/alm-picker.
Given the task (speech / sound / music / multi-audio / long-audio, output modality, latency, license), output:
- Model. Qwen2.5-Omni-7B · Qwen3-Omni · SALMONN · Audio Flamingo 3 · AF-Next · LTU · GAMA · Gemini 2.5 Pro (API) · GPT-4o Audio (API). One-sentence reason.
- Benchmark subset to validate. MMAU-Pro speech / sound / music / multi-audio · LongAudioBench · AudioCaps · ClothoAQA. Pick the axis that matches the user task.
- Output modality. Text-only · text + speech (Qwen-Omni, GPT-4o Audio). Budget for an additional speech decoder if needed.
- Guardrails. Reject prompts that require multi-audio comparison when your model's multi-audio score is < 30% (near-random). Diarize before LALM for > 10-minute inputs.
- Escalation. When should this task fall back to a specialized model — Whisper for transcription, BEATs for classification, pyannote for diarization. LALM is not the best of each.
Refuse to ship multi-audio comparison tasks without verifying your model scores > 40% on the MMAU-Pro multi-audio subset. Refuse long-audio (> 10 min) without upstream diarization. Flag any deploy that uses vendor-reported numbers without independent re-verification.
Example input: "Compliance audit: transcribe 10-minute bank-call recordings + detect if the agent read the mandatory disclosure."
Example output:
- Model: Whisper-large-v3-turbo for transcription + Gemini 2.5 Pro (via API) for disclosure-check QA over the transcript. LALM direct on raw audio is tempting but long-audio LALM accuracy drops past 10 min.
- Benchmark subset: MMAU-Pro speech subset (Gemini 2.5 Pro = 73.4%) — covers the speech-reasoning axis. Also spot-check on your own 50-call gold set.
- Output modality: text-only. Speech output not needed for an audit report.
- Guardrails: diarize with pyannote 3.1 first; send per-speaker segments separately; log confidence score per call.
- Escalation: if a call fails the disclosure check, route to human reviewer instead of autonomous flag.
Security
- Security scanPassed, 95 of 100Pattern checks on the skill's files.
- Model reviewPassed
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingskills/alm-picker/SKILL.md
Popularity
- 7-day trend score
- 5
Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/watcher-hermes/hermes-skills/alm-picker)