菜单
Analyze audio file for music video prompt scheduling. Picks the right analyzer (ffmpeg-only for quick energy/structure, librosa for BPM/key/F0/JSON-export-for-LLM), and outputs a prompt schedule template with timestamps. Also works on generated audio for diagnostic comparison.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Canonical first-pass when a ComfyUI-AudioLoopHelper workflow won't run or fails validation. Use when the user mentions "workflow won't run", "dependency cycle detected", "validation failed", "Failed to convert ... to INT", "control_after_generate", or describes a recent ComfyUI render that didn't get past prompt validation, or reports a loop/full-song render that is pathologically slow or re-encodes (text-encode / audio / keyframe) every iteration. Runs the canonical diagnosis (ComfyUI log tail, audit, DAG inspection, exec log review, launch-flag check for --cache-none) BEFORE any inline orjson scripting. Reference: docs/reference/debug_tools.md.
Structural diff of two ComfyUI workflow JSON files — node types, widget values, link topology, subgraph schemas — with noise filtered out (link IDs, node positions, ordering). Use when debugging "why does workflow A behave differently than B" or verifying a migration script produced the intended shape change.
Generate an LTX 2.3 prompt that makes a character speak SPECIFIC dialogue, formatted per the official LTX 2.3 prompt guide (short quoted segments with acting directions between, accent, acoustic environment). Use when the user wants to put words in a character's mouth for the AV inversion / one-pass / keyframe workflows (voice-cloned dubbing, dialogue replacement).
Scaffold a new scripts/apply_*.py workflow migration script from scripts/templates/, enforcing the project's idempotence + paired-audit-check convention. Use when adding a new fix that mutates example_workflows/*.json or stages an experimental variant.
Dispatch three parallel Explore agents to validate a structural refactor BEFORE writing changes. Use when the user proposes restructuring multiple CLAUDE.md / docs / agent / skill files (signal: ≥ 3 files, ≥ 2 different concerns), or when planning a curation pass on documentation. Cost ~30s of agent time; turns vague "compress this" into a concrete edit list and prevents premature subtree splits.
Run the AudioLoopHelper test suite with correct uv groups, rootdir, and ComfyUI loader verification
| name | audio-analyze |
| description | Analyze audio file for music video prompt scheduling. Picks the right analyzer (ffmpeg-only for quick energy/structure, librosa for BPM/key/F0/JSON-export-for-LLM), and outputs a prompt schedule template with timestamps. Also works on generated audio for diagnostic comparison. |
| disable-model-invocation | true |
Last updated: 2026-04-30
Analyze an audio file (input track OR generated output) and surface what's needed for prompt scheduling, or for diagnosing differences between source and generated audio.
/audio-analyze <audio_file> [--trim <seconds>] [--rich]
| Mode | Tool | When |
|---|---|---|
| Quick energy + structure (no librosa) | scripts/analyze_audio.py | Triage; works on a fresh clone with zero Python deps |
| Rich features (BPM, key, F0, JSON for LLM) | scripts/analyze_audio_features.py | Authoritative analysis for prompt-schedule generation; default when --rich is passed or when the user asks for BPM/key/structure-with-confidence |
If the user gave an audio file with no further context, run the quick analyzer first; if the output suggests vocal+music but no prompt schedule will be built, stop there. If they're going to author prompts (or diagnose generated audio), run the rich analyzer.
uv run scripts/analyze_audio.py <audio_file> --trim <seconds>
--rich)uv run --group analysis scripts/analyze_audio_features.py <audio_file> \
--scene-diversity --montage --style \
--json /tmp/audio_features.json
--trim <seconds> if the user gave one. The flags pull style hints +
montage cuts + section diversity into the JSON for LLM consumption./tmp/audio_features.json, surface BPM, key, F0 trajectory, structure.scripts/spectrogram_to_reference.py --audio <wav> makes the audio
reviewable in the chat.wave — that gives
misleading values).--group analysis on the uv invocation.ffmpeg -i <mp4> -vn -acodec pcm_s16le <wav> first.docs/guides/audio_analysis_guide.mddocs/guides/prompt_workflow_end_to_end.mdscripts/spectrogram_to_reference.py