pharmscribe
niceloading0127/pharmscribe
Turn pharmacology video lectures into bilingual (EN/中文) Word notes + a searchable, source-traceable knowledge base. Vendor-neutral CLI & Claude Code skill.
Instalação
npx convez-ai niceloading0127/pharmscribe1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
PharmScribe
Turn pharmacology / pharmaceutical-science video lectures into bilingual (English term + Chinese explanation) Word study notes, and accumulate every lecture into a local, queryable knowledge base with exact source traceability (jump back to the precise video + timestamp).
Vendor-neutral. You do not need a specific AI membership. Downloading, transcription, the knowledge base, and Word generation are plain programs. Only the "extract knowledge from the transcript" step uses an AI model — and that model is swappable: Claude, OpenAI (ChatGPT), Gemini, a free local model via Ollama, or none at all.
Two ways to use it
A. Standalone CLI (works for everyone)
# Pharmacology lecture, using a free local model (no API key, no membership)
pharmscribe ingest "https://youtu.be/XXXX" --engine local
# Using your own API key from any one vendor
pharmscribe ingest "https://youtu.be/XXXX" --engine openai # ChatGPT/OpenAI
pharmscribe ingest "https://youtu.be/XXXX" --engine gemini # Google Gemini
pharmscribe ingest "https://youtu.be/XXXX" --engine claude # Anthropic
# Cross-query your whole library (no AI needed)
pharmscribe query --target CYP3A4
pharmscribe query --drug warfarin --category interaction
pharmscribe videos
The engine is chosen with --engine or the PHARMSCRIBE_ENGINE env var.
B. Claude Code skill (for Claude users)
Drop the folder in ~/.claude/skills/ and just ask Claude to summarize a video.
In this mode Claude does the extraction itself (and can additionally see the
video frames — slides, mechanism diagrams, structures — for richer notes).
Choosing an engine
--engine | Needs | Membership? | Notes |
|---|---|---|---|
local | Ollama + a model (ollama pull qwen2.5) | None | Free, private, offline |
openai | pip install openai, OPENAI_API_KEY | Your own key | |
gemini | pip install google-generativeai, GEMINI_API_KEY | Your own key | |
claude | pip install anthropic, ANTHROPIC_API_KEY | Your own key | |
none | nothing | None | Offline keyword fallback; low quality |
Only the engine you pick needs installing.
Requirements
yt-dlp,ffmpegon PATHpip install python-docxsqlite3(Python standard library)- An engine from the table above (or
--engine none) - Whisper (
pip install openai-whisper) — optional, only when a video lacks captions
Output
- A bilingual Word document per lecture, grouped by pharmacology category (mechanism / PK / PD / ADR / interactions / SAR / indications), with clickable "re-watch @ timestamp" links (and embedded frames in skill mode).
- A growing SQLite knowledge base queryable by drug, target, pathway, topic, or free text — combinable as AND filters, every result traced to its source.
Notes & limits
- The portable CLI extracts from captions/transcript (text), so it runs identically across all engines. Frame-aware (multimodal) extraction currently requires the Claude skill path, where the model can view images directly.
- Educational note-taking only — not medical or clinical advice.
Roadmap
- Chemical structure → SMILES (optical structure recognition)
- Anki / spaced-repetition export straight from the knowledge base
- Frame-aware extraction for the CLI engines as their vision APIs allow
License
MIT — see LICENSE.