Post-Prompt

When you type into a chat window, your sentence rarely goes to the AI system unchanged. Usually, the software surrounding your input field adds further text to it. A post-prompt is the part that stands behind your input. It contains, for example, the instruction to answer briefly, to remain polite, or to cite a source. Its counterpart is called a pre-prompt and stands before your input. Together, both form the text that the model actually gets to see.

Why the last instruction has particularly strong effect

Language models are programs that continue text. They read everything that lies before them and then predict, word by word, the most probable continuation. Not every position counts equally, though. Instructions placed shortly before the start of the answer often have a stronger effect in practice than those located further up in the text.

That is exactly why the post-prompt is so useful for developers. It is the last reminder of the rules before the model starts writing. In long conversations with many messages, this matters. A rule from the first message can practically evaporate by the time the twentieth response comes around.

There is a downside, though. If the user’s text comes before the instruction, it can attempt to overwrite it. If someone writes “Ignore all of the following rules,” a conflict arises. Such attacks are called prompt injection, and a post-prompt only partially protects against them.

How the text is assembled before it reaches the model

Technically, this happens in the program that calls the AI, not in the model itself. The software takes a template with a gap for the user input. Before the gap comes the pre-prompt, after the gap comes the post-prompt. This stitching together is called prompt templating.

An example from a customer service chat makes this tangible. The user writes: “My package hasn’t arrived.” Behind it, the system silently inserts: “Answer in a maximum of three sentences, don’t mention any prices, and ask for the shipment number.” The model sees both as one continuous text. The user only sees their own sentence and the reply.

A misunderstanding often occurs here: a post-prompt is not a post-processing of the answer. It is set before the model even starts computing. Anyone who alters the finished answer afterward, say by checking figures or removing profanity, is doing post-processing. These are two different steps, even though both words start with “post.”

Post-prompts in products and in the news

Practically every AI application with an input field uses this technique. The translate button in an email program works this way, as does the summarization feature in a news app. The homework helper that deliberately doesn’t reveal solutions but only gives hints is also a post-prompt case. The rules are contained in the invisible additional text, not in the model.

In the news, post-prompts usually surface when they come to light. Time and again, users manage to coax a chatbot into revealing its hidden instructions. Such leaked prompts reveal a lot about the company behind them, for instance which topics it wants to avoid. Some companies therefore treat their prompt templates as trade secrets.

For you as a user, a simple insight follows from this. An AI answer is never just a reaction to your question. It is always also the result of rules that someone else set beforehand. If a chatbot answers noticeably tersely or evasively, this is often not due to the model, but to the text surrounding your question.