Prompting as a Method: Principles, PE Parameters, Design Languages, and RAG/FT

Prompt Engineering fundamentals

Prompt engineering (PE) improves instruction clarity and enables more useful model outputs. Different models require different prompt formats, and official guidance recommends testing multiple cases to find the best fit for a task.

Core elements

Mnemonic: Be specific , Work in steps , Provide context .

RGC principle: Role , Result , Goal , Context , Constraint .

The IBM CHI 2024 paper “Challenges and Opportunities for Responsible Prompting” defines a prompt anatomy and lists resources (books, courses, datasets, GitHub repos).

Design patterns and parameters

Classification of design patterns

A survey (arXiv 2402.07927) classifies PE by application domain. The taxonomy distinguishes basic prompt modes—zero‑shot, few‑shot, instruction fine‑tuning—and advanced techniques such as Chain‑of‑Thought (CoT), Self‑Consistency, ReAct, and multi‑agent frameworks.

Parameters for non‑reasoning models

model : higher‑performance models are more expensive and have higher latency.

temperature : larger values increase randomness/creativity; for factual tasks set to 0. Modify either temperature or top_p, not both.

top_p : nucleus sampling; e.g., top_p=0.1 limits token selection to the top 10 % probability mass.

max_completion_tokens : hard token limit; the model usually stops earlier based on a stop sequence.

stop : a list of stop sequences that abort generation.

See the OpenAI API reference for full details.

Prompt format preferences

Claude models – XML tags (the model was trained on large XML corpora).

OpenAI latest models – Markdown format (observed in system prompts of GPT‑4.5 and GPT‑4o).

Tool‑calling scenarios – JSON schemas to describe function signatures, parameter types, and required fields.

Simple tasks (e.g., DALL‑E system prompt) – plain text.

LISP – niche format occasionally used in Chinese AI communities.

Prompt Engineering vs Retrieval‑Augmented Generation vs Fine‑tuning

RAG

Addresses three major LLM limitations: knowledge timeliness, private‑data gaps, and hallucinations.

Simple RAG = vector database + LLM. Drawbacks: partial‑similarity recall, semantic loss, lingering hallucinations.

Advanced RAG extends the context window and improves retrieval accuracy (e.g., self‑RAG, GraphRAG, qwen‑Agent).

Fine‑tuning (FT)

Identify high‑impact workflow nodes, collect high‑quality datasets, and run targeted FT projects.

Benefits: handle more examples, use shorter prompts, protect private data, and deploy cheaper specialised models.

Typical FT pipeline:

Prepare training data (often JSONL) with structured {"prompt": ..., "completion": ...} records.

Choose a training method: Supervised Fine‑Tuning (SFT), Direct Preference Optimization (DPO), or Reinforcement Fine‑Tuning (RFT).

Run the training job on a pre‑trained model that already performs well on broad tasks.

Comparative strengths (Vibe‑coding landing‑page example)

PE – lowest cost, fastest iteration, but weaker consistency on long dialogs and limited domain knowledge.

RAG – higher factual accuracy and better adherence to design guidelines; slower response and may be overly conservative.

FT – best domain‑specific performance and fastest response; highest maintenance and development cost.

Evaluation methods

PE evaluation: human or AI grading against predefined criteria (relevance, correctness, style).

RAG evaluation: retrieval‑quality metrics (precision, recall), hallucination detection, and how much retrieved knowledge is used.

FT evaluation: task‑specific scores such as BLEU, exact‑match rate, and assessment of generalisation on out‑of‑distribution data.

Andrew Ng advises building evaluation infrastructure early to avoid technical debt.

Illustrative examples and code snippets

Prompt‑style correction (❌ → ✅):

❌:
Summarize the text below as a bullet point list of the most important points.

{text input here}

✅:
Summarize the text below as a bullet point list of the most important points.

Text: """
{text input here}
"""

Zero‑shot vs few‑shot keyword extraction:

✅ Zero‑shot:
Extract keywords from the below text.

Text: {text}

✅ Few‑shot:
Text 1: Stripe provides APIs that web developers can use to integrate payment processing into their websites and mobile applications.
Keywords 1: Stripe, payment processing, APIs, web developers, websites, mobile applications

Text 2: OpenAI has trained cutting‑edge language models that are very good at understanding and generating text.
Keywords 2: OpenAI, language models, text understanding, text generation

JSON function schema (Manus system prompt example):

{
  "type": "function",
  "function": {
    "name": "[tool]_[action]",
    "description": "[tool description]. Use for [scenario].",
    "parameters": {
      "type": "object",
      "properties": {
        "[param1]": {"type": "[type]", "description": "[description]"},
        "[param2]": {"type": "[type]", "description": "[description]", "enum": ["option1", "option2"]}
      },
      "required": ["[required1]", "[required2]"]
    }
  }
}

LISP‑style prompt template:

(prompt-design
  (role-definition
    (:name "Domain expert")
    (:domain "Machine learning"))

  (task-description
    (:goal "Explain neural network principles")
    (:audience "Non‑technical users"))

  (constraints
    (:depth-level :intermediate)
    (:taboo-list "mathematical formulas"))

  (output-spec
    (:format "metaphor explanation")
    (:length 500)))

Future of prompting

OpenAI’s CEO Sam Altman notes that as LLM capabilities increase, the need for intricate prompt engineering will diminish. Prompting will remain useful but become simpler: fewer explicit constraints, more high‑level guidance, and a shift toward philosophical direction rather than detailed rule‑crafting.