# Prompt Guide ### Introduction Prompt engineering is the practice of designing structured inputs that guide large language models (LLMs) to generate accurate, relevant, and context-aware outputs. Rather than relying solely on model size or training data, well-crafted prompts significantly influence performance across tasks such as summarization, content generation, analysis, and decision support. Inspired by industry best practices from sources such as IBM Think and applied research in generative AI, this document introduces a **Prompting Cookbook** tailored for **FPT AI Marketplace**, providing reusable prompting techniques applicable across diverse models and use cases. Throughout this document, selected examples reference specific models available on FPT AI Marketplace (such as DeepSeek-V3.2-Special) strictly for illustration. The described prompting techniques remain fully applicable across all supported models. *** ### Understanding Prompt Structure A prompt typically consists of four logical components: | Component | Description | | ------------------ | ------------------------------------------------------------------- | | Role / Context | Defines who the model should act as or the situation it operates in | | Task Instruction | Clearly states what the model is expected to do | | Constraints | Specifies tone, format, length, or rules | | Output Expectation | Describes the desired structure of the response | Well-structured prompts improve consistency, interpretability, and portability across models. *** ### 1. Zero-Shot Prompting **Use case:** Baseline tasks, fast answers, exploratory queries Zero-shot prompting asks the model to perform a task without providing examples, relying entirely on its pretrained knowledge. **Template** ``` Explain {topic} in a clear and simple way for {target audience}. ``` **Example** ``` Explain cloud-native architecture in simple terms for non-technical business users. ``` **🔍 Illustrative Example (Model: DeepSeek-V3.2-Special)** You are using DeepSeek-V3.2-Special on FPT AI Marketplace. Explain the role of a Business Analyst\ in an AI product development lifecycle.\ Provide a concise, structured explanation. > This example demonstrates zero-shot prompting behavior on a reasoning-capable model.\ > The same prompt structure can be reused across other Marketplace models. *** ### 2. Chain-of-Thought Prompting **Use case:** Reasoning-heavy explanations, structured analysis, consulting-style responses Chain-of-Thought (CoT) prompting encourages the model to reason step by step. **Template** ``` Step 1: Identify the problem Step 2: Analyze key factors Step 3: Provide a conclusion Now answer the following: {question} ``` **Example** ``` Step 1: Identify the business problem Step 2: Analyze constraints Step 3: Recommend an AI solution Question: How should enterprises adopt generative AI responsibly? ``` **🔍 Illustrative Example (Model: DeepSeek-V3.2-Special)** You are using DeepSeek-V3.2-Special on FPT AI Marketplace. Step 1: Identify challenges in adopting AI platforms\ Step 2: Analyze organizational and technical risks\ Step 3: Summarize best practices Question: What are key considerations when enterprises adopt AI marketplaces? > This example illustrates step-by-step reasoning.\ > CoT prompting is model-independent and can be applied to other reasoning-capable models. *** ### 3. Prompt Chaining **Use case:** Multi-step workflows, document pipelines, agent-like behaviors Prompt chaining decomposes a complex task into smaller, sequential prompts. **Example flow** 1. Generate outline 2. Expand sections 3. Refine tone and format **Example** ``` Prompt 1: Create an outline for an AI governance guideline. Prompt 2: Expand each section with examples. Prompt 3: Rewrite for executive stakeholders. ``` **🔍 Illustrative Example (Model: DeepSeek-V3.2-Special)** You are using DeepSeek-V3.2-Special on FPT AI Marketplace. Prompt 1: Create an outline for onboarding AI models onto a marketplace.\ Prompt 2: Expand each step with validation and compliance details.\ Prompt 3: Rewrite the content for a business audience. > This example demonstrates how prompt chaining enables workflow-style generation.\ > The same chaining logic applies across all Marketplace models. *** ### 4. RAG-Aware Prompting **Use case:** Enterprise knowledge grounding, factual accuracy, policy-driven responses RAG-aware prompting instructs the model to ground its responses in retrieved or provided knowledge. **Template** ``` Using the provided context below, answer the question accurately. If information is missing, state that explicitly. Context: {retrieved_content} Question: {question} ``` **Example** ``` Context: Internal policy document on data residency. Question: What data residency rules apply to AI workloads? ``` **🔍 Illustrative Example (Model: DeepSeek-V3.2-Special)** You are using DeepSeek-V3.2-Special on FPT AI Marketplace. Using the provided onboarding documentation,\ explain compliance requirements for publishing AI models.\ If details are unavailable, explicitly state limitations. > This example highlights enterprise-safe prompting patterns.\ > RAG-aware prompting is recommended for all factual and compliance-sensitive use cases. *** ### Best Practices for FPT AI Marketplace * Keep prompts clear, modular, and reusable * Avoid coupling prompts too tightly to one model * Use examples only when output consistency is required * Prefer structured outputs for automation and integration * Iterate prompts based on observed outputs *** ### Appendix — Model-Specific Examples (Illustrative) The following examples demonstrate how selected prompting techniques\ may be applied to specific models available on FPT AI Marketplace.\ These examples are non-exhaustive and provided for reference only. #### Example: DeepSeek-V3.2-Special (Text Reasoning) **Use case:** Enterprise explanation with structured reasoning Explain the AI model onboarding process on FPT AI Marketplace.\ Structure the answer into: * Prerequisites * Validation steps * Deployment readiness This example highlights structured reasoning behavior and does not imply\ model exclusivity or technique dependency. ### 5. Applied Use Case: Education Summarization & Evaluation Use case: Automated grading support, step-by-step logic analysis, structured feedback. This recipe demonstrates how to combine Direct Instructions, Role Prompting, and Task-Specific Output to assist educators. The goal is to digest a student's lengthy math solution into a concise summary with an accuracy assessment. **The Configuration (Recipe)** This use case utilizes specific parameters to balance mathematical precision with natural language feedback. * Model: `gpt-oss-20b` (or similar reasoning-capable models) * Temperature: `0.6` (Balances accuracy with natural phrasing) * Max Tokens: `600` (Ensures the full summary and evaluation are generated without cutoff) * Reasoning Effort: `medium` (Optimizes latency to \~5-10 seconds while maintaining logical depth) **The Prompt Structure** System Prompt: Plaintext ``` Reasoning: medium You are a mathematics education expert. Your task is to summarize a student's math solution. REQUIREMENTS: 1. Identify the solution method used. 2. List key steps (2-3 lines per step). 3. Evaluate the accuracy of the solution. 4. Identify strengths and weaknesses. 5. Provide improvement suggestions (if any). OUTPUT FORMAT: - Method: [Name of method] - Steps: [Concise summary] - Evaluation: [Correct/Incorrect] - Strengths: [Specific details] - Improvements: [Suggestions] ``` User Prompt: Plaintext ``` Summarize the following student math solution: Problem: Solve the system of equations: { 2x + 3y = 13 ; 3x - 2y = 4 } Student Solution: Step 1: I will use the elimination method... [...Insert full student text here...] ``` **🔍 Illustrative Output (Actual Response)** The model processes the input (\~912 tokens) and generates a structured evaluation (\~536 tokens): > * Method: Elimination method > * Steps: > 1. Multiplied equations to align coefficients of y (6 and -6). > 2. Added equations to eliminate y, solving for x = 38/13. > 3. Substituted x back into the original equation to find y = 31/13. > 4. Verified results by substituting x and y into both original equations. > * Evaluation: Correct > * Strengths: > * Used an appropriate method and correctly performed multiplication/addition. > * Detailed calculation avoiding errors. > * Full verification of roots. > * Improvements: > * Avoid converting to decimals prematurely; keep fractions for higher precision. > * Multiplication steps could be optimized to avoid repetition. This example illustrates how Task-Specific Instructions can transform raw text into standardized data suitable for educational platforms. *** ### References * IBM Think – Prompt Engineering Techniques\ * IBM Research – Prompting and Reasoning Methods\ [https://research.ibm.com](https://research.ibm.com/) * DeepSeek Documentation\ * General Prompt Engineering Guide\ --- # Agent Instructions: Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter: ``` GET https://ai-docs.fptcloud.com/api-reference/ai-marketplace/prompt-guide.md?ask= ``` The question should be specific, self-contained, and written in natural language. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.