Prompt

When you will meet it

You meet it on your very first API call and in the first prompt-engineering article you read. Most people assume the prompt is the sentence they typed; inside an agent, that sentence is often a small slice of what the model receives. Miss this and you will keep rewriting your own sentence while the model is actually looking at a large pile of other content the harness stuffed in.

Minimal example

// 你在聊天框只打了一句「幫我修這個 bug」
// 但送進模型的 prompt 實際長這樣（簡化示意）：
[
  { "role": "system",    "content": "你是寫程式的助手，只改工作區內的檔案…" },
  { "role": "user",      "content": "（幾輪之前的對話…）" },
  { "role": "assistant", "content": "（模型的舊回答…）" },
  { "role": "user",      "content": "工具清單與相關檔案內容…\n幫我修這個 bug" }
]

Look at the structure: the model receives one long role-tagged sequence in which your words are the final sliver. So "tuning the prompt" in practice often means tuning the whole assembly — and the harness, not you, owns the assembly rules.

What people get wrong

• Assuming the prompt is just your last sentence. In agent settings the system prompt, tool definitions and history dominate; when behaviour looks wrong, find a way to see the actual assembled request before editing anything.
• Treating prompting as talking to a person, expecting a sincere or forceful tone to shift the model's capability boundary. A prompt steers what the model continues with; it does not change what the model can do.