watcher-hermes/hermes-skills/any-to-any-pipeline-auditor
Claude Code and agent skill by watcher-hermes/hermes-skills
- community
Checking that the SKILL.md file is still in the repository.
Audit a conversational any-to-any design and compute the latency budget for a MIO / AnyGPT / Moshi-family stack.
Install
npx vskill@latest install watcher-hermes/hermes-skills/any-to-any-pipeline-auditorSKILL.md
The same SKILL.md was published earlier as watcher-hermes/hermes-skills/any-to-any-pipeline-auditor.
Given a conversational product (speech in / speech out, optional vision, optional music), a model size, and a target latency, audit the any-to-any design and produce a viable configuration.
Produce:
- Modality mix. Which modalities in, which out. Pick family: MIO / AnyGPT (discrete tokens, 4 modalities), Moshi (speech+text focused, inner monologue), Unified-IO 2 (vision-rich).
- Shared vocabulary plan. ID ranges for text + image + speech + music + separators. Total size typically 40-50k.
- Tokenizer stack. BPE + SEED + SpeechTokenizer-RVQ + Encodec. Highlight which are still bottlenecks (speech quality typically).
- Training curriculum. Four-stage MIO recipe, or two-stage for speech-focused Moshi.
- TTFAB latency budget. Mic encoder + prefill + first token + residual decode + speech decoder. Compare to ~500ms conversational bar.
- Quality-vs-latency pareto. Smaller model for low latency, larger for higher quality; rough numbers per A100/H100.
Hard rejects:
- Proposing separate models per modality when the requirement is conversational fluidity. The pipeline latency stacks and feels worse.
- Using a speech tokenizer with only 1 codebook layer. Quality will be robotic for any production voice.
- Claiming MIO's TTFAB matches GPT-4o. It does not yet; Moshi 160ms is the closest open number.
Refusal rules:
- If target TTFAB <200ms, refuse MIO-scale (8B+) and recommend Moshi-class (7B, tuned for speech) or a smaller speech-specialized model.
- If user wants studio-quality voice output, refuse open residual-VQ and recommend ElevenLabs / chained-TTS until open quality catches up (Qwen3-Omni / Moshi2).
- If user wants image generation during a voice call, refuse streaming-speech-first and propose a split pipeline with mode-switching.
Output: one-page audit with modality mix, vocab plan, tokenizer stack, curriculum, TTFAB latency, quality-latency pareto. End with arXiv 2409.17692 (MIO), 2410.00037 (Moshi), 2402.12226 (AnyGPT).
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/any-to-any-pipeline-auditor/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/any-to-any-pipeline-auditor)