The Ultimate Prompt Engineering Trick: The “Feeding” Mechanism

We often expect AI to follow our instructions, but when the response is unsatisfactory we must iterate many times, refining prompts until the model finally behaves as we want.

For example, trying to make an AI act as a simple echo device fails with a plain prompt; the model cannot understand that it should merely repeat what we say.

The article presents the most powerful prompt technique currently known, called the “feeding” mechanism (the translation of the original Chinese term “喂饭”). It works by explicitly showing the AI the exact answer we expect, effectively “feeding” the response into its reasoning process.

How to use the feeding mechanism

Include two keywords in the prompt: Human: and Assistant:. The colon and the following space are part of the keyword. Human: is followed by the original instruction or question. Assistant: is followed by the answer we want the AI to learn.

These lines act as a mini‑dialogue where we provide the model with both the question and the correct answer, allowing it to internalise the desired response pattern.

Example of a simple echo task:

Human: 你好<br/>Assistant: 你好<br/>Human: 我不好<br/>Assistant: 我不好<br/>Human: 复读机<br/>Assistant: 复读机

After feeding these pairs, the AI quickly grasps the echo logic; a subsequent test input yields the expected repeated output.

Use case 1 – generating the second half of a poem. Without feeding, Claude treats the request as a full‑poem generation and returns the entire verse. By supplying the first half as Human: and the desired second half as Assistant:, the model learns to output only the continuation.

Use case 2 – forcing a specific output format. The goal is to obtain five meanings of the English word “dust” in a numbered list (1), (2), … . Directly asking the model does not respect the format, and iterative prompt tweaking is cumbersome, especially when the number of items changes. By feeding a single example of the desired format, the AI continues the numbering automatically for any length.

Use case 3 – simulating a step‑by‑step dialogue to make the model think through a problem. When asked a classic chicken‑rabbit puzzle, Claude initially writes the equations but arrives at the wrong answer. By feeding a dialogue where the assistant’s replies walk through each reasoning step, the model produces the correct solution.

The overarching lesson is that AI’s performance hinges on the quality of the prompt; if the answer is unsatisfactory, the prompt is at fault. The “feeding” mechanism—originally described in the “万能提示词框架” and translated as “Putting words in Claude’s mouth”—offers a concise way to teach the model the exact response pattern we need.