Versioning API Prompts in Code

When you build a standard React app, you version control your code using Git. If you break the CSS, you can look at the Git history and revert to the previous commit.

When you build an AI application, your System Prompts are your code. A one-word change to a prompt can drastically alter the behavior of your Agent.

The Problem with String Literals

Many beginners build AI apps by hardcoding massive strings directly inside their API routes:

// BAD PRACTICE
app.post('/api/chat', async (req, res) => {
  const response = await anthropic.messages.create({
    system: "You are an AI that helps users. Always output JSON...",
    messages: req.body.messages
  });
});

If you do this, your routes.js file will quickly become 5,000 lines long, filled with messy paragraphs of English text mixed in with your JavaScript logic.

Extracting Prompts to Files

To maintain a professional codebase, you should extract your prompts into their own dedicated files, usually ending in .txt or .md.

/src
  /api
    routes.js
  /prompts
    /v1
      customer_service_agent.txt
      qa_agent.txt

In your routes.js, you simply read the file:

const fs = require('fs');

app.post('/api/chat', async (req, res) => {
  // Read the prompt from the file system
  const systemPrompt = fs.readFileSync('./prompts/v1/customer_service_agent.txt', 'utf8');

  const response = await anthropic.messages.create({
    system: systemPrompt,
    messages: req.body.messages
  });
});

Why Versioning Matters

When you separate your prompts into files, they become first-class citizens in your Git repository.

If your QA Agent suddenly starts approving buggy code, you don't have to guess what changed. You can run git diff on qa_agent.txt and see exactly which word was added or deleted by your coworker yesterday.

Furthermore, you can easily implement A/B testing. You can create customer_service_agent_v2.txt. You can route 50% of your users to v1 and 50% to v2, and monitor the Evals pipeline to see which prompt performs better in production.

Treat Prompt Engineering like Software Engineering. Version your prompts.