Master Advanced Prompt Engineering: Boost LLM Performance with Proven Techniques

In recent months large models such as ChatGPT and GPT‑4 have been released, demonstrating strong emergent abilities, yet their outputs can be inconsistent; a key factor is prompt design.

System Messages

System messages appear at the beginning of a prompt to give the model context, define what it should or should not answer, and specify the desired response format.

Typical system messages can instruct the assistant to answer only using provided context, to admit uncertainty, or to perform specific tasks such as extracting entities in JSON.

Assistant is a large language model trained by OpenAI.

It should answer questions using only the given context and say "I don't know" if unsure.

It can help users with tax‑related queries.

It can extract entities from text and return them as a JSON object.

Example output format:

{
  "name": "",
  "company": "",
  "phone_number": ""
}

Even well‑designed system messages may be contradicted by the model.

Few‑Shot Learning

Few‑shot learning supplies a small set of examples within the prompt to give the model additional context for a new task.

These examples guide the model to respond in a desired way, simulate specific behavior, or provide seed answers for common questions.

Non‑Conversation Scenarios

Large models can also be used outside chat, e.g., for sentiment analysis.

Use Explicit Instructions

Prompt order matters: start with the task description, then provide context or examples. This improves output quality.

Repeat Instructions at the End

Because later prompt tokens can have a stronger influence, repeating the instruction at the end can help ensure compliance.

Operations on the Output

Appending specific words or phrases (e.g., "Here’s a bulleted list of key points:\n- ") can steer the model to produce the desired format.

Add Syntax

Using punctuation, headings, or separators like "---" helps the model parse the prompt and can serve as a stop condition.

Decompose Tasks

Breaking a task into smaller steps generally improves LLM performance.

Chain‑of‑Thought Prompting

Ask the model to reason step‑by‑step, presenting each intermediate step, which reduces errors and makes evaluation easier.

Provide Real Context

Supplying authentic data reduces the model’s workload and error chance; the closer the raw data to the final answer, the better.

The Microsoft guide also contains additional prompt‑engineering tricks for further exploration.