What is a prompt template, and how should you version and test prompts in production?

A prompt template is a prompt with gaps. The fixed part, the instructions and examples, is written once; the gaps, such as $email or {question}, are filled in for each request. Every real LLM app works this way, because the same instructions go with thousands of different inputs.

Keep prompts in files, not in strings

A prompt buried in a Python string is easy to change by accident and hard to review. Kept as its own file, it is a first-class part of the code:

• Versioned. Each meaningful change is a new version, triage-v2.md beside triage-v1.md, or a new commit, so you can say which prompt produced an answer and roll back a bad change.
• Logged. Record the prompt's name, version and a hash of its text with every call. When an answer goes wrong a month later, you know exactly what the model was told.
• Reviewed. A change to a prompt is a change to behaviour, and deserves the same review as a change to code.
• Tested. A small set of cases with known right answers runs against every version before it ships, and grows into an eval setEval SetThe list of test cases an LLM app is graded on after every change. Each case is an input and what a good answer must do, so a score going up or down shows whether the change helped..

How to fill the gaps safely

• Fail on a missing value. A template library that silently leaves $email blank produces a call that looks fine and answers nothing. Python's string.Template.substitute raises an error instead.
• Mark where outside text goes. Put the filled-in value inside tags, such as <email> and </email>, so the model can tell your instructions from a customer's words. It helps the model read the prompt correctly; it does not by itself stop prompt injectionPrompt InjectionAn attack that hides instructions in text a model reads, such as a web page, an email or a file, so the model follows the attacker instead of you., so keep the limits on what the model can do in code.
• Choose a placeholder style that will not clash. Curly braces collide with JSON examples in the prompt; $name placeholders, or a library like Jinja, avoid that.

Put the fixed part first

The fixed part of a template should come before the variable part, ideally in the system promptSystem PromptInstructions the builder of an AI app gives the model before you type anything. It sets the role, the tone and the rules, and it outranks what users say. with the user's input last. That order is what prompt cachingPrompt CachingLetting the API keep the processed start of a prompt for a few minutes, so later requests that begin the same way are cheaper and quicker. needs: the cache covers the prompt from the start up to a marked point, so a stable beginning can be reused across requests, while a timestamp or customer name near the top would make every request look new. On current Claude models a prefix needs at least 512 tokens to be cached at all.