🧠 Personal Prompt Router (Experiment)

A weekend experiment to build a lightweight AI control plane that:

1. Classifies incoming prompts
2. Routes them to a model
3. Toggles reasoning mode
4. Logs every decision
5. Analyzes behavior

---

🚀 Setup

1️⃣ Install dependencies

npm install

2️⃣ Configure environment variables

Copy the example file:

cp .env.example .env

Then edit .env and set your OpenRouter API key:

OPENROUTER_API_KEY=your_openrouter_api_key_here
OPENROUTER_BASE_URL=https://openrouter.ai/api/v1
CLASSIFIER_MODEL=openrouter/openai/gpt-5.2-chat
EXECUTION_MODEL_DEFAULT=openrouter/openai/gpt-5.2-chat

⚠️ Never commit your real .env file.

---

🧪 Running the Router

Start the CLI:

npm run dev

You’ll be prompted to enter a prompt.

The system will:

• Classify it
• Show routing decision
• Execute via OpenRouter
• Log the decision in logs/

---

🧪 Run Full Test Suite (Automated)

To benchmark the router against the 15 categorized prompts:

tsx src/runTestSuite.ts

(Optional) Add a script in package.json:

"test-suite": "tsx src/runTestSuite.ts"

Then run:

npm run test-suite

This will:

• Parse prompts from test-suite.md
• Classify each prompt
• Execute via the router
• Measure latency
• Compute overall and per-category accuracy

---

📊 Analyze Logs

After running multiple prompts:

tsx src/analyzeLogs.ts

This prints:

• Total runs
• Average latency
• Category distribution
• Reasoning usage breakdown

---

🗂 Project Structure

src/
  classifierPrompt.ts   → Classification instructions
  routingTable.ts       → Static routing rules
  openrouter.ts         → OpenRouter API integration
  logger.ts             → JSONL logging
  analyzeLogs.ts        → Analytics script
  index.ts              → CLI entrypoint

test-suite.md           → 15-prompt evaluation set
logs/                   → Router decision logs

---

🎯 Experiment Goals

• ≥80% classification accuracy
• Measure reasoning usage
• Measure latency impact
• Evaluate subjective output quality