How ly-xxx/ace-taffy-skill works
永雏塔菲.skill combines an instruction-driven persona skill for Codex/Claude Code with a research pipeline that collects public Weibo/Bilibili material and prepares speech and style data. The primary user workflow is to invoke the skill with a writing or roleplay request; the agent uses the skill definition and persona references to produce a response. The secondary workflow refreshes public sources, transcribes selected video audio, audits transcripts, and builds corpus/training artifacts. README documents additional style-analysis capabilities; unsampled tool wiring is not assumed.
Main entry points: transcripts Transcript files · build_corpus_public.py Corpus builder · download_bilibili_media.py Media download · raw Raw captures
Components 25
Persona Runtime
- Skill instructions
SKILL.md - Persona profile
persona.md - Skill metadata
meta.json - Task prompts
prompts
Public Sources
- Weibo collection
tools/collect_weibo_public.py - Bilibili collection
tools/collect_bilibili_public.py - Raw captures
sources/raw
Speech Data
- Media download
tools/download_bilibili_media.py - Downloaded media
sources/media - Audio transcription
tools/transcribe_audio.py - Transcript files
sources/transcripts - Batch STT
tools/batch_bilibili_stt.py - Transcript audit
tools/audit_transcripts.py - Audit artifacts
sources/processed - Training selection
tools/build_training_set.py - Training selections
sources/processed/training
Style Knowledge
- Persona references
references - Corpus builder
tools/build_corpus_public.py - Processed corpus
sources/processed/corpus - Style bank builder
tools/build_style_bank.py - Speech-pattern analysis
tools/analyze_meow_pattern.py
Other components
- Skill user
- Codex / Claude Code
- Bilibili
Connections 25
- Skill user → Codex / Claude Code: requests output
- Codex / Claude Code → Skill instructions: loads instructions
- Codex / Claude Code → Persona profile: reads persona
- Codex / Claude Code → Skill metadata: reads metadata
- Codex / Claude Code → Task prompts: uses prompts
- Codex / Claude Code → Persona references: consults guidance
- Codex / Claude Code → Skill user: returns draft
- Weibo collection → Weibo: fetches public posts
- Bilibili collection → Bilibili: fetches public data
- Weibo collection → Raw captures: writes captures
- Bilibili collection → Raw captures: writes captures
- Media download → Bilibili: downloads streams
- Media download → Downloaded media: writes media
- Batch STT → Media download: runs downloader
- Batch STT → Audio transcription: runs transcription
- Audio transcription → Transcript files: writes transcripts
- Transcript audit → Transcript files: reads transcripts
- Transcript audit → Audit artifacts: writes audit results
- Training selection → Audit artifacts: reads audit data
- Training selection → Training selections: writes selections
- Corpus builder → Raw captures: reads captures
- Corpus builder → Transcript files: reads transcripts
- Corpus builder → Processed corpus: writes corpus
- Style bank builder → Processed corpus: reads corpus data
- Speech-pattern analysis → Transcript files: analyzes speech