missuo/koe/koe-setup

Koe Setup & Configuration

Guide the user through Koe's initial setup and ongoing configuration. This skill helps with installation, credential setup, personalized dictionary generation, system prompt customization, and hotkey configuration.

Configuration Files

All config files live in ~/.koe/ and are auto-generated on first launch:

~/.koe/
├── config.yaml          # Main configuration (ASR, LLM, feedback, hotkey)
├── dictionary.txt       # User dictionary (hotwords + LLM correction terms)
├── system_prompt.txt    # LLM system prompt (correction rules)
└── user_prompt.txt      # LLM user prompt template (usually no need to change)

Workflow

Step 1: Check Installation Status

Check if Koe is installed and if ~/.koe/ exists:

# Check if config directory exists
ls -la ~/.koe/
 
# Check if Koe.app is installed
ls /Applications/Koe.app 2>/dev/null || mdfind "kGMDItemDisplayName == 'Koe'" -onlyin /Applications -onlyin ~/Applications

If not installed, guide the user:

# Option A: Homebrew (recommended)
brew install koe --cask
 
# Option B: Download from GitHub Releases
# https://github.com/missuo/koe/releases/latest

After installation, launch Koe once to generate default config files, then grant three permissions in System Settings > Privacy & Security:

1. Microphone — for audio capture
2. Accessibility — for auto-paste (Cmd+V simulation)
3. Input Monitoring — for global hotkey detection

Step 2: Configure ASR Credentials

Koe uses Doubao (豆包) ASR by ByteDance. The user needs Volcengine (火山引擎) credentials.

Ask the user: "Do you already have Volcengine (火山引擎) ASR credentials? If not, I can guide you through getting them."

Guide to get credentials:

1. Go to https://console.volcengine.com/speech/app
2. Create an app, then copy App ID (app_key) and Access Token (access_key)

Then update ~/.koe/config.yaml:

asr:
  app_key: "<their App ID>"
  access_key: "<their Access Token>"

Step 3: Configure LLM

Koe supports any OpenAI-compatible API for text correction. Ask the user which LLM provider they use.

Common configurations:

# OpenAI
llm:
  base_url: "https://api.openai.com/v1"
  api_key: "sk-..."  # or "${OPENAI_API_KEY}"
  model: "gpt-4o-mini"
 
# Any OpenAI-compatible endpoint
llm:
  base_url: "https://your-provider.com/v1"
  api_key: "..."
  model: "your-model"

Recommend fast, cheap models since latency matters for voice input (gpt-4o-mini, deepseek-chat, etc.).

Important: The api_key field supports environment variable substitution with ${VAR_NAME} syntax. If the user prefers not to put API keys in plain text config files, suggest using environment variables.

Step 4: Personalized Dictionary

This is the most impactful customization. Ask the user about their profession and domain:

"What's your primary work domain? For example: frontend development, backend/DevOps, data science, iOS development, product management, academic research, etc. I'll generate a tailored dictionary for you."

Based on their answer, generate a comprehensive dictionary file at ~/.koe/dictionary.txt. The dictionary serves two purposes:

1. ASR hotwords — improves speech recognition accuracy for these terms
2. LLM correction context — the LLM prioritizes these spellings

Guidelines for dictionary generation:

• One term per line, lines starting with # are comments
• Include proper capitalization (e.g., PostgreSQL not postgresql)
• Include brand names, tools, frameworks, libraries relevant to their domain
• Include common abbreviations and acronyms they'd use verbally
• Include team/product-specific terms if the user mentions them
• Group terms by category with comment headers for readability
• Aim for 100-300 terms for a good starting dictionary

Example structure:

# Programming Languages
TypeScript
JavaScript
Python
Rust
 
# Frameworks & Libraries
React
Next.js
Tailwind CSS
 
# Cloud & Infrastructure
Cloudflare
AWS
Kubernetes
Docker
 
# Tools
VS Code
GitHub Actions
Terraform
 
# Domain-Specific Terms
# (add terms specific to user's work)

After generating, remind the user they can always add more terms later — it's just a text file.

Step 5: System Prompt Customization

The default system prompt is tuned for software developers working in mixed Chinese-English. Based on the user's profession, consider adjusting:

Read the current system prompt:

cat ~/.koe/system_prompt.txt

Possible customizations based on profession:

• Non-technical users: Remove or simplify the technical term rules, add rules for their domain terminology
• English-only users: Remove Chinese-specific rules (spacing, punctuation), focus on English conventions
• Academic/research: Add rules for citation terms, paper-specific jargon, LaTeX terms
• Medical/legal: Add domain-specific formatting rules

Only modify the system prompt if the user's needs clearly differ from the defaults. The default prompt works well for most developer use cases.

Important: The user_prompt.txt template usually does not need changes. Only modify it if the user has specific needs for the prompt structure.

Step 6: Hotkey Configuration

By default, Koe uses the Fn key. If the user wants a different trigger key, explain the available options:

hotkey:
  # Options: fn | left_option | right_option | left_command | right_command
  trigger_key: "fn"

Changes take effect automatically within ~3 seconds (no restart needed).

Step 7: Feedback Sounds

If the user wants to disable sound effects:

feedback:
  start_sound: false   # Silence on recording start
  stop_sound: false     # Silence on recording stop
  error_sound: true     # Keep error sounds (recommended)

Verification

After setup, guide the user to test:

1. Launch Koe (or it may already be running in the menu bar)
2. Open any text input (Notes, browser, terminal)
3. Press and hold the trigger key, speak a sentence, release
4. Or tap the trigger key once to start, speak, tap again to stop
5. The corrected text should be pasted automatically

If something doesn't work, check:

• Menu bar icon should show Koe is running
• All three permissions are granted in System Settings
• ASR credentials are correct (check app_key and access_key)
• LLM endpoint is reachable and API key is valid