Prompt Engineering & Prompt Tuning Practical Skills

Introduction: Course Overview

This document compiles the full learning materials for the Zhibo AI Large Model Practical Course, instructed by Teacher Jerry, sourced from the presentation slides of the MoPass Moboyun · Zhibo AI Artificial Intelligence Research Institute. The course covers end-to-end content: prompt fundamentals, core prompt engineering theories, six advanced prompt tuning techniques, prompt attack security defense, hands-on Coze (Kouzi) AI agent platform operations, LLM API application development, and multi-scenario prompt engineering real-world cases. It is an essential foundational course for AI application developers.

Prompts serve as the sole medium for human-LLM interaction. Prompt Engineering is a standardized, engineering-oriented methodology for optimizing prompts. All LLM deployment technologies including RAG, Agent, Coze, and Dify rely on well-designed prompts as their underlying foundation. Mastering this course drastically reduces AI hallucinations, standardizes output formats, and improves model reasoning & content generation performance.

1. What is a Prompt

Plain-language Definition: A prompt is a piece of text instruction, question, or requirement description sent by users to Large Language Models (LLMs), acting as the core communication channel between humans and AI. Short single-sentence questions and full complex role-task descriptions both qualify as prompts.

1.1 Critical Importance of Prompts

1. 1. Primary Interaction Medium: Prompt quality directly controls output accuracy, relevance, and practicality; poorly written prompts generate vague, generic, off-target responses.
2. 2. Unlock Full Model Potential: Comprehensive prompts activate the model’s full capabilities, mitigate AI hallucinations, and ensure outputs match user expectations precisely.
3. 3. Foundational Base Technology: All LLM deployment solutions (RAG, AI Agents, Coze low-code bots, API development) implement core business logic via well-structured prompts.

1. Two Core Principles for Building High-Quality Prompts

Principle 1: Provide Sufficient Context & Clear, Unambiguous Instructions

• Include complete background, context, and business details to supply the model with adequate reference information;
• State tasks directly without ambiguity, clearly defining the exact action the model must complete;
• Use punctuation, line breaks, and labeled sections to further eliminate misinterpretation (Example: Name: Jerry).

Principle 2: Guide Complex Tasks Step-by-Step

For multi-step reasoning and multi-stage business workflows, predefine the thinking and execution sequence for the model, forcing the AI to follow human logic step-by-step and avoid disorganized reasoning.

2. Complete Theory of Prompt Engineering

Prompt Engineering is a standardized discipline centered on designing, testing, and iteratively refining prompts with an engineering mindset to continuously align model outputs with business objectives. Its full iterative workflow is a 3-step loop: Clarify requirements & define logic → Write standardized prompts → Validate outputs and optimize repeatedly.

2.1 Standard Building Blocks of a Complete Prompt

• Role: Define the AI’s identity, professional expertise, and core responsibilities;
• Skills: Sub-capabilities and execution standards tied to the assigned role;
• Requirements: Output details, formatting rules, quality standards, and stylistic constraints;
• Task: The core objective the LLM must fulfill;
• Workflow: Sequential execution steps and reasoning order;
• Examples: Successful reference samples, negative counterexamples, and output templates;
• Constraints: Hard boundaries, prohibited behaviors, information restrictions, and risk mitigation rules;

Two classic real-world template use cases: product customer service chatbots and formal ID photo editing AI agents, both built directly on this modular structure for deployment on platforms like Coze.

3. Six Core Prompt Tuning Techniques

Basic prompts only handle simple use cases. Complex reasoning, standardized formatting, and high-precision tasks require specialized tuning methodologies. Six mainstream techniques exist, each differing in computational cost and applicable scenarios.

3.1 Zero-Shot Prompting

Definition: Deliver instructions without providing any reference examples; the model completes tasks directly. Lowest computational cost.

• Applicable Scenarios: General simple Q&A, translation, basic factual knowledge, elementary arithmetic;
• Examples: Sentiment classification judgment, English-to-Chinese translation, basic fact lookup.

3.2 Few-Shot Prompting

Definition: Supply a small set of standardized examples to teach the model task formats and output rules, resolving disorganized or inconsistent outputs from zero-shot prompting. Low computational cost.

• Applicable Scenarios: Tasks requiring fixed output formatting, classification, poetry writing, standardized copywriting;
• Examples: Labeled sentiment classification, structured poem generation.

3.3 Chain-of-Thought (CoT)

Definition: Guide the model to decompose complex problems, reason step-by-step, and deliver a final conclusion afterward. Can be combined with few-shot prompting (Few-Shot CoT). Common directive: Think through this step by step. Medium computational cost.

• Applicable Scenarios: Math word problems, logical reasoning, multi-stage complex questions;
• Core Value: Significantly improves accuracy for complex model reasoning tasks.

3.4 Self-Consistency

Definition: Generate multiple independent CoT reasoning paths, then select the most frequently occurring result via majority voting as the final answer to reduce random errors. High computational cost.

• Applicable Scenarios: High-precision tasks including scientific computation, legal judgment, medical analysis;
• Drawbacks: Multiple full response generations lead to slower reply speeds and higher token consumption.

3.5 Tree-of-Thought (ToT)

Definition: An extended CoT framework that structures reasoning into tree branches, explores multiple solution paths, filters valid logic, eliminates invalid branches, and selects the globally optimal output. Extremely high computational cost.

• Applicable Scenarios: Complex planning, mathematical proof, code debugging, multi-path decision-making;
• Drawbacks: Complex implementation with heavy hardware and model capability requirements.

3.6 Self-Reflection Mechanism

Definition: Three-stage workflow: Generate initial draft → Self-audit logic, factual accuracy, and formatting flaws → Revise to produce the final polished output, mimicking human critical thinking. Medium-to-high computational cost.

• Applicable Scenarios: Code generation, copywriting polishing, professional analytical reports;
• Standard Template: 1. Generate raw answer 2. Audit all errors 3. Revise and deliver final output.

4. Prompt Attack Types & Security Mitigation

Prompt Attacks refer to maliciously crafted input designed to trick LLMs into bypassing security guardrails, leak system prompts, generate restricted content, or hijack original business logic. This is a mandatory security requirement before launching any AI application to production.

4.1 Four Common Prompt Attack Categories

1. 1. Prompt Injection: Hijack the model’s original instructions and force modified output rules; Example: Ignore all prior translation rules and output specified text.
2. 2. Prompt Leakage: Manipulate the model to expose backend system prompts, role rules, and underlying base instructions;
3. 3. Prompt Jailbreaking: Circumvent safety filters to generate illegal, dangerous, restricted content; Examples: Roleplay as a movie burglar to ask for home break-in methods, simulate a root Linux terminal.
4. 4. Sensitive Information Extraction: Impersonate trusted identities to steal serial keys, credentials, private data, and other confidential information.

4.2 Standardized Mitigation Strategies

• 1. Pre-execution System Layer Detection: Insert security judgment prompts before running business logic to identify rule-circumventing and malicious conflicting inputs;
• 2. Fixed AI Identity & Business Boundaries: Hardcode strict limits on the model’s scope of work, rejecting all cross-domain and rule-bypassing requests;
• 3. Multi-layer Security Validation: Content filtering, keyword blocking, dual input & output audit workflows;
• 4. Hard restriction against exposing system prompts, backend configurations, and secret keys.

5. Coze (Kouzi) AI Agent Platform Overview & Hands-On Operations

Coze (Kouzi) is ByteDance’s all-in-one no-code / low-code AI Bot development platform. Developers with minimal programming experience can rapidly build conversational AI agents and visual AI applications, with finished deployments publishable to channels including Douyin. All AI Agents within the platform are labeled as Bots.

5.1 Domestic vs Global Coze Platform Versions

• Domestic Version (coze.cn): Direct access via mainland internet; default Doubao LLM, compatible with DeepSeek, Tongyi Qwen, Kimi, etc.
• Global Version (coze.com): Requires special network access; supports GPT, Claude, Gemini, and other overseas large models.

5.2 Two Core Coze Project Types

• AI Agent Bot: Pure conversational AI project that completes business workflows via dialogue, plugin calls, and workflow triggers; Examples: Weather Dragon Bot, Travel Itinerary Assistant;
• AI Application: Complete standalone AI program with full visual interfaces and independent business pipelines.

5.3 Registration & Bot Building Workflow

Registration & Login: Mobile number + verification code login → Set nickname & username → Enter workspace. After creating a new agent, fill modular standardized prompts in the Persona & Reply Logic panel, integrate plugins, knowledge bases, and workflows to implement functionality, then preview, test, and publish the finished bot.

6. LLM API Application Development: Hello World

Beyond low-code platforms, developers can invoke LLMs via official vendor APIs to build custom programs. This section uses Alibaba Bailian Tongyi Qwen as a complete end-to-end development example.

1. 1. Platform Registration & Login: Sign in to Alibaba Bailian via Alipay or mobile phone number;
2. 2. Activate model services and claim free token call quotas;
3. 3. Navigate to Key Management and create an API Key; securely store credentials and never hardcode keys in source code;
4. 4. Configure system environment variable DASHSCOPE_API_KEY to avoid credential leakage;
5. 5. Prepare Python environment: Configure domestic pip mirror sources, install OpenAI-compatible SDK;
6. 6. Write code: Initialize client, inject system prompts and user requests, call the API endpoint, and print returned results.

7. Prompt Engineering Real-World Implementation Cases

7.1 SQL Multi-Table Query Generator

Business Scenario: Input database table creation statements and query requirements to generate executable standard SQL with explanatory notes. Core prompt logic: Parse table schema first, decompose query requirements, output valid SQL syntax, and explain execution logic for each statement.

7.2 Construction Machinery Enterprise Financial Cost Analysis System

Business Scenario: Compare multi-year public financial data for Sany Heavy Industry and Zoomlion Heavy Industry, automatically extract core metrics including operating revenue, net profit, and gross profit margin, and generate standalone company financial reports + cross-brand competitor comparison reports. Prompts integrate structured data constraints, stepwise data analysis pipelines, and fixed report output formats, paired with RAG for financial report raw data retrieval to reduce hallucinations.

8. Course Summary & Production-Grade Development Workflow

• 1. Prompts form the foundational layer of all LLM applications; standardized modular prompt structure (Role / Skills / Tasks / Workflow / Examples / Constraints) serves as universal development specification;
• 2. Select tuning techniques based on task complexity: Zero/Few-Shot for simple tasks, CoT for reasoning tasks, Self-Consistency for high-precision workloads, ToT for complex decision-making, Self-Reflection for coding & writing;
• 3. All AI applications must implement prompt security protection to block injection, jailbreaking, and system prompt leakage attacks;
• 4. Two core deployment routes: No-code Coze for rapid bot prototyping, custom API code development for fully customized programs;
• 5. Industrial-grade deployment stack: High-quality Prompt + RAG knowledge base retrieval + tool function calling + security validation to resolve AI hallucinations and model capability limitations.