Master Prompt Engineering: A Universal Framework for Building Effective LLM Prompts

Prompt Universal Framework

The ultimate goal for developers is to reduce complexity by applying a single, reusable method to most problems; in the era of large models, that method is Prompt engineering. The quality of a Prompt directly determines model performance.

Many online articles list isolated Prompt tricks without a coherent structure, leaving readers unsure where to start. Drawing on Tencent engineers' practical experience, this guide organizes Prompt engineering into a systematic workflow that helps anyone become a Prompt engineer.

1. Prompt Universal Framework

We split a Prompt into four components: role + problem + goal + requirements . This template yields a "passable" Prompt for the first version of any task.

2. Framework Refinement

2.1 Define Role

A role acts as an "ability package" for the model, similar to a job description in recruitment. By specifying identity, education, experience, and relevant abilities, the model knows which capabilities to activate.

Example role template:

You are a skilled {{desired identity}} with {{desired education level}} and {{years of experience}}. Your task is {{task related to the problem}} and you possess the following abilities: {{required abilities}}.

If you lack domain knowledge, you can borrow information from job postings to fill the template (e.g., a financial analyst role for a finance‑analysis task).

2.2 Describe Problem & Set Goal

The problem description carries the most information. Clearly state what you want the model to do and the expected outcome. For complex tasks, decompose the problem into sub‑steps (similar to Chain‑of‑Thought) before prompting.

2.3 Add Requirements

Place constraints at the end of the Prompt so the model, which performs text completion, gives higher weight to the most recent context. Use explicit markers (e.g., "#", "<>", "```", "[]", "-") or Markdown syntax to structure the Prompt.

3. Enhancing Prompts with RAG and Few‑Shot

3.1 Retrieval‑Augmented Generation (RAG)

RAG combines semantic retrieval (embedding + vector database) with the model to inject up‑to‑date knowledge. Popular frameworks include LangChain, Milvus, LlamaIndex, and Pinecone. When vector databases are unavailable, traditional keyword matching can still provide useful retrieval.

3.2 Few‑Shot Examples

Providing a few high‑quality examples (few‑shot) improves accuracy without full fine‑tuning. To avoid over‑reliance on examples, keep them concise and representative, and use RAG to retrieve the most relevant examples dynamically.

3.3 Memory Management

Short‑term memory refers to the current conversation context, while long‑term memory stores historical interactions. Injecting relevant history into the Prompt helps the model maintain continuity across sessions.

3.4 Handling Specialized Domains

Large models excel at general knowledge but need domain‑specific information for specialized tasks (e.g., test‑case generation). Collect and structure domain knowledge, then use RAG to retrieve it on demand, turning static expertise into dynamic, model‑usable input.

4. Advanced Techniques

4.1 Controlling Model Determinism

Two key parameters affect output randomness: Temperature (controls diversity) and Top‑P (nucleus sampling). Lower temperature yields more deterministic, higher‑quality text; higher temperature increases creativity. Top‑P limits the cumulative probability mass of considered tokens.

4.2 Automatic Prompt Optimization

Algorithms such as APE (Automatic Prompt Engineer), APO, and OPRO use the model itself to generate candidate Prompts, score them on a validation set, and iteratively refine them via Monte‑Carlo search, gradient descent, or language‑model‑as‑optimizer techniques. These methods can discover more effective Prompt phrasing (e.g., improving "Let's think step by step" to "Take a deep breath and work on this problem step‑by‑step").

5. Summary

By following the four‑stage workflow— Prompt Framework → Framework Refinement → Adding More Information (RAG, Few‑Shot, Memory) → Advanced Techniques (parameter tuning, automatic optimization) —developers can construct robust, reusable Prompts for any task, manage Prompt libraries as structured tables, and continuously improve model performance.