ChatGPT 0720 Update: Custom Instructions, System Messages, and Implementation Guide

During a late‑night test of the ChatGPT website, the 0720 update was discovered. The update brings two main changes for Plus users: the GPT‑4 usage limit is doubled from 25 to 50 requests per three hours, and a new "Custom Instructions" feature is added, allowing users to define personal roles and desired answer styles.

Custom Instructions, translated as "personalized directives," stem from the concept of instruction tuning used in large language model training (e.g., InstructGPT). An instruction is a structured prompt that guides the model’s behavior, while a prompt is more of a reference.

To try the feature, Plus users open the Beta features panel and find the "Custom instructions" menu. The interface asks for two parts: a personal profile (location, work, hobbies, topics of interest, goals) and a response style (formality, length, address, opinion stance).

Enabling Custom Instructions changes the model’s output dramatically. Without the feature, a query about integrating unit tests in a front‑end project receives a generic answer. With a profile set as "front‑end engineer, Typescript, code solution," the model suggests Jest and provides a full installation and configuration guide.

Technically, Custom Instructions are equivalent to injecting a user‑defined prompt at the start of each conversation. This eliminates the need to repeat the same context for every new chat. OpenAI’s official feedback highlights the inconvenience of re‑entering fixed prompts and the value of role‑based definitions.

The article also discusses implementing similar functionality via system messages in the GPT API. The messages array can contain a system role with high weight, followed by user and assistant messages. A typical system message might include three sections: AssistantProfile, UserProfile, and AssistantReplyStyle.

[
  {
    "role": "system",
    "content": "#AssistantProfile: JD interview officer\n#UserProfile: Front‑end intern\n#AssistantReplyStyle: proactive, serious, comprehensive"
  },
  {
    "role": "user",
    "content": "Hello interviewer, I am ready."
  }
]

In the HiBox plugin, a new configuration key HiBox.config.chatgptProfiles allows multiple profiles to be defined, each with fields for assistant profile, user profile, and answer style. When a request is made, the profile is read, converted to English (to increase weight), and inserted into the system message.

export function getSystemMessageWithProfile() {
  // Read the user‑defined custom instruction (Chinese template for readability)
  const profileStr = GlobalState.get('chatgptCurrentProfile');
  // Convert to English (English template has higher weight)
  return profileStr
    .replace('[系统简介]', '#AssistantProfile')
    .replace(' [个人简介]', '\n#UserProfile')
    .replace(' [回答风格]', '\n#AssistantReplyStyle');
}

// When calling GPT
const body = {
  // ... other fields ...
  systemMessage: getSystemMessageWithProfile() || 'You are ChatGPT, a large language model trained by OpenAI. Please answer as concisely as possible.',
};

Local tests in HiBox show that using the custom instruction and system‑message approach yields expected results for translation and interview simulation scenarios.