umeboshi

Umeboshi — AI-Powered Anki Card Generator for Medical Students

日本語版 | English version

Turn lecture slides and textbooks into high-quality Anki flashcards with bilingual annotations, visual aids, and cloze deletions — in 2 commands.

git clone https://github.com/Ark-Karim/umeboshi.git && cd umeboshi && ./install.sh
claude
#    Inside Claude Code: /anki-card-generator-without-hs

No API key required. Claude Code runs on your existing Claude Pro ($20/mo) or Max ($100–200/mo) subscription. Transcription uses your machine's GPU (local Whisper, English default).

---

What This Does

Each lecture becomes a deck of structured Anki cards through a 6-step pipeline:

1. Transcribe — Lecture audio → text (local Whisper, English; pluggable backends)
2. Structure — Key concepts → organized sections with exam-relevant highlights
3. Cloze — Smart fill-in-the-blank cards with contextual grouping
4. Annotate — English medical terms annotated with furigana (ruby), drug names color-coded with mnemonics
5. Visualize — Auto-generated diagrams, tables, flowcharts, and comparison charts
6. Import — Cards pushed directly to Anki via AnkiConnect (port 8765)

Output: Each card includes a ~3,000-character bilingual detailed description, exam-focused high-yield summary, visual aids, and one-by-one cloze progression.

---

Real-World Usage

Here's how a pharmacology student turns a lecture script into a reviewed Anki deck:

Step 1: Generate Cards

cd C:\Users\chiba\StudySpace\pharmacology2Final
claude
# Run: /anki-card-generator-without-hs

The skill will prompt for:

• Working directory — where your script file lives (e.g., C:\Users\chiba\StudySpace\pharmacology2Final)
• Script file — your lecture text (e.g., script/20.txt)

Sub-agents then transcribe → structure → cloze → annotate → visualize → import. Generation takes ~3–10 minutes per lecture depending on length and model.

Pro tip — direct sub-agents for better quality: When the skill asks for additional instructions, you can tell sub-agents to:

• Emphasize exam-relevant content and bold high-yield points
• Target detailedDescription at ~3,000 characters (2,000–4,500 range)
• Prioritize topics the professor flagged for the exam

Step 2: Manual Cloze Refinement

After generation, manually review the cards in Anki's browser (b key). Add or adjust cloze deletions ({{c1::...}}) to match your study style.

Step 3: Import with Bilingual + Color Annotations

# In Claude Code:
/anki-card-importer-without-hs

Provide:

• Working directory (e.g., C:\Users\chiba\StudySpace\pharmacology2Final)
• Original filename without extension (e.g., 17)
• Deck name (e.g., active::pharmacology2::17)

Pro tip — color-coded drug mnemonics: Tell the D3 sub-agent:

Assign a theme color to each drug class. Use <span style="color: #xxxxxx;"> on drug names and stems. Pick colors from JIS common color names that act as mnemonics (e.g., EP1 antagonist → "Epipen" → yellow #e6b422). Document why you chose each color on first occurrence.

This makes drug names visually distinct and reinforces memorization through color association.

---

Customization Tips

What	How	Why
Exam focus	Tell the generate sub-agent which topics are high-yield	Cards surface what's tested
Detail depth	Set detailedDescription character target (recommended: 2,500–3,500)	Balances context vs. review speed
Color mnemonics	Use <span style="color: ...;"> for drug/stem names	Visual memory hook for pharmacology
Bold key points	Request bolding of exam-critical content in sub-agent instructions	Quick scan during review
Difficulty level	Pass level is easy/hard to the import sub-agent	Controls card complexity
Non-English transcription	Switch to Fireworks/OpenRouter backend (see Alternative Paths)	Japanese, Chinese, etc.

---

Prerequisites

• Node.js 18+ and Git
• Claude Code CLI (npm install -g @anthropic-ai/claude-code)
• Anki 24.06+ with AnkiConnect enabled
• Anki running in the background (cards are pushed via AnkiConnect on port 8765)

---

Detailed Setup

1. Clone and Install

git clone https://github.com/Ark-Karim/umeboshi.git
cd umeboshi
./install.sh

2. Import the Umeboshi-Kaname Note Type

1. Open Anki

2. Tools → Manage Note Types → Add

3. Create a new note type named "Umeboshi-Kaname"

4. Switch to the Cards... view and copy-paste:

   • notetype/Umeboshi-Kaname/Front.html → Front template tab
   • notetype/Umeboshi-Kaname/Back.html → Back template tab
   • notetype/Umeboshi-Kaname/Styling.css → Styling tab

5. Switch to the Fields... view and add these fields (order matters):

   #	Field name
   1	text
   2	extra
   3	imagesExtra
   4	detailedDescription
   5	highYieldSummary
   6	visualAids
   7	visualAidsExtra
   8	additionalResources
   9	oneByOne
   10	umeboshiNoteId

   The full field definition is in notetype/Umeboshi-Kaname/notetype.json.

3. Start Generating

cd /path/to/your/lecture/materials
claude
# Inside Claude Code:
/anki-card-generator-without-hs

---

Alternative Paths

DeepSeek (no Claude subscription)

DeepSeek's API lets you use Claude Code without an Anthropic subscription.

# Linux/macOS
export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN="<your DeepSeek API Key>"
export ANTHROPIC_MODEL=deepseek-v4-pro[1m]
export ANTHROPIC_DEFAULT_OPUS_MODEL=deepseek-v4-pro[1m]
export ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-v4-pro[1m]
export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-v4-flash
export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4-flash

# Windows PowerShell
$env:ANTHROPIC_BASE_URL="https://api.deepseek.com/anthropic"
$env:ANTHROPIC_AUTH_TOKEN="<your DeepSeek API Key>"
$env:ANTHROPIC_MODEL="deepseek-v4-pro[1m]"
$env:ANTHROPIC_DEFAULT_OPUS_MODEL="deepseek-v4-pro[1m]"
$env:ANTHROPIC_DEFAULT_SONNET_MODEL="deepseek-v4-pro[1m]"
$env:ANTHROPIC_DEFAULT_HAIKU_MODEL="deepseek-v4-flash"
$env:CLAUDE_CODE_SUBAGENT_MODEL="deepseek-v4-flash"

Get your API key from platform.deepseek.com/api_keys. Cost: ~$0.30 per 90-minute lecture (Japanese).

OpenRouter (alternative provider)

OpenRouter provides multi-provider failover and budget controls.

export ANTHROPIC_BASE_URL=https://openrouter.ai/api/v1
export ANTHROPIC_AUTH_TOKEN="<your OpenRouter API Key>"
# Optional: /fast mode
export ANTHROPIC_DEFAULT_HAIKU_MODEL=anthropic/claude-haiku-4-5
export CLAUDE_CODE_SUBAGENT_MODEL=anthropic/claude-haiku-4-5

Get your key from openrouter.ai/keys.

External Transcription (Japanese / non-English)

export FIREWORKS_API_KEY="fw_..."    # https://console.fireworks.ai → API Keys
# or
export OPENROUTER_API_KEY="sk-or-..."  # https://openrouter.ai/keys

Then edit plugins/registry.json to switch the active transcription backend:

{ "backends": { "transcribe": { "active": "<your-plugin-name>" } } }

See docs/PLUGIN_DEVELOPMENT.md to create a custom backend. Cost: ~$0.004/min audio (varies by provider).

Switching Backends

Umeboshi uses a plugin system. Edit plugins/registry.json to swap backends:

{
  "backends": {
    "transcribe": { "active": "local-whisper" },
    "generate":   { "active": "claude-code" },
    "import":     { "active": "ankiconnect" }
  }
}

See docs/ARCHITECTURE.md and docs/PLUGIN_DEVELOPMENT.md.

---

Optional: Cloud Sync

Synchronize card Q&A across devices via Cloudflare Workers:

1. npm install -g wrangler
2. cd claudeflare && npx wrangler kv:namespace create UMEBOSHI_KV
3. Update wrangler.toml with the KV namespace ID
4. npx wrangler deploy
5. In Anki: press s → ClaudeFlare URL → paste your worker URL

See plugins/sync/claudeflare/SKILL.md for details. Disabled by default.

---

Optional: Build Note Type from Source

cd notetype/Umeboshi-Kaname
npm install ejs
./build

Compiles src/front.ejs + src/components/*.ejs → Front.html.

---

Cost

Path	Cost	Notes
Default (Claude subscription + local Whisper)	$0 beyond subscription	Claude Pro $20/mo or Max $100–200/mo
DeepSeek (alt generation)	~$0.30 / 90-min lecture	Requires DEEPSEEK_API_KEY
External transcription (alt)	~$0.004/min audio	Fireworks/OpenRouter, for non-English

---

Troubleshooting

Problem	Solution
Cards not appearing	Anki must be running with AnkiConnect enabled (port 8765)
ClaudeFlare not configured	Set worker URL in Anki Settings (press s)
Transcription fails (external)	Verify API key is set and valid
Card generation fails (DeepSeek)	Verify DEEPSEEK_API_KEY env var
Skills not found	Re-run ./install.sh
Local Whisper missing	pip install openai-whisper or brew install whisper-cpp

---

Project Structure

plugins/         Plugin system — swappable backends
skills/          Claude Code skill definitions (the pipeline)
notetype/        Anki card template (Umeboshi-Kaname) + source EJS
claudeflare/     Cloudflare Worker for cloud sync (optional)
docs/            Architecture, plugin development
examples/        Sample content and source attribution

---

License

GNU AGPL-3.0 — see LICENSE. Includes code derived from AnKing Note Types — see NOTICE.

Community

AI-generated cards may contain errors — verify medical content against trusted sources.