A Comprehensive Guide to Designing Structured Prompts for AI Agents

Core Elements of Structured Prompts

Role

# Role\nYou are a [Xiaohongshu viral copywriting expert] specializing in [beauty/skincare]. Your tone is friendly, like a best friend, and you excel at using emojis and exclamation marks to stir emotions.

Context/Background

# Background\nBelow is a recent excerpt (published Dec 2025) of the IT equipment management and information security policy of XX Technology Company.

Skills/Tools

# Skills\n1. Call the weather‑query plugin to obtain real‑time weather data.\n2. Parse the city name from the user query (supports Chinese and English).\n3. Extract special weather‑alert information from the knowledge base.

Constraints

# Constraints\n- Answer only weather‑related questions.\n- Responses must include temperature, weather condition, and air‑quality index.\n- Keep the answer concise and avoid verbosity.

Workflow

# Workflow\n1. Receive user input and identify the question type.\n2. If weather data is needed, call the weather‑query plugin.\n3. If special weather alerts are required, retrieve information from the knowledge base.\n4. Consolidate information and generate a response according to the format.

Output

# Output\nProvide a concise Chinese reply containing:\n- Current temperature (°C)\n- Weather condition (e.g., sunny, cloudy, rain)\n- Air‑quality index (AQI) and level\n- Suggested clothing thickness or outdoor precautions

Advantages of Structured Prompts

Clear logic and easy maintenance – Modular design makes responsibilities explicit, supporting team collaboration and version rollback. Coze workflow version‑control enables safe iteration.

Strong control, stable output – Explicit constraints and workflow guide the model, reducing hallucinations. Reported relevance improvement: 30‑40%.

High reusability, efficiency boost – Standardized templates can be reused across projects; Coze statistics show average developer efficiency increase of 2.3×.

Major Prompt Frameworks

R‑C‑S‑W‑O Model (Role, Context, Skills, Workflow, Output)

# Role\nYou are a [XX‑domain expert] proficient in [specific skills], communicating in [specific style].\n\n# Context\nHere is detailed information about [task background]: [background description].\n\n# Skills\n1. [Skill 1 description]\n2. [Skill 2 description]\n3. [Skill 3 description]\n\n# Workflow\n1. [Step 1]\n2. [Step 2]\n3. [Step 3]\n\n# Output\nPlease use [format] and [language style] to provide a response of [content depth].

Applicable scenarios:

Complex tasks requiring multiple plugins (e.g., travel planning, data analysis)

Multi‑step reasoning workflows (e.g., code optimization, legal consulting)

Enterprise‑level applications (e.g., IT policy Q&A, internal knowledge‑base management)

Case study – travel planning agent on Coze:

Role: Professional travel planner

Context: User plans a family trip from Beijing in summer 2026 with a budget of 15,000 CNY

Skills: Weather query, attraction recommendation, hotel booking, flight price comparison plugins

Workflow: Determine destination → query weather → recommend attractions → arrange accommodation and transport

Output: List‑style daily itinerary with accommodation, dining suggestions and precautions

Zeng Yingjie Three‑Stage Model

# Who am I: [identity/background]\n# What am I doing: [core task/goal]\n# What are my requirements: [specific details/constraints/style]

Applicable scenarios: rapid prototyping, simple single‑turn tasks, beginner developers.

Case study – weight‑loss plan generator:

# Who am I: I am a fitness coach with 10 years of fat‑loss experience.\n# What am I doing: Provide a one‑week diet and exercise plan.\n# What are my requirements: Include breakfast, lunch, dinner, and snack suggestions; home‑friendly exercises; friendly tone; each suggestion ≤30 words; total calories 1,500‑1,800 kcal.

CRISPE Framework (Capacity & Role, Insight, Statement, Personality, Experiment)

# Capacity and Role\nYou are an AI assistant proficient in Chinese‑English academic writing, familiar with journal submission requirements.\n\n# Insight\nUser is drafting a paper on "Large Language Models in Medical Diagnosis" and needs abstract polishing.\n\n# Statement\nRewrite the abstract to meet IEEE Transactions on Biomedical Engineering standards, limit to 250 words, use passive voice and academic terminology.\n\n# Personality\nMaintain a professional, rigorous academic tone; avoid colloquial language; ensure logical clarity and sufficient argumentation.\n\n# Experiment\nInput example: "We developed a GPT‑4‑based medical diagnosis system that accurately identifies common diseases."\nOutput example: "This study proposes a GPT‑4‑based diagnostic system that achieves high‑precision identification of common diseases, offering a novel technical pathway for clinical decision‑making."

Applicable scenarios: precise output style control, cross‑domain knowledge integration, long‑term projects requiring iterative optimization.

Five Key Prompt‑Optimization Techniques

1. Clarify Context and Goal

Providing sufficient, relevant background and defining clear objectives raises relevance by ~35% (research). Use frameworks such as CRISPE or R‑C‑S‑W‑O.

Original: "Help me write a speech about AI."\nStructured: "You are an AI technology expert preparing a 30‑minute keynote for the 2026 World AI Conference. Topic: 'Generative AI in Healthcare'. Include: 1) latest technical advances; 2) real‑world cases in diagnosis and drug discovery; 3) current technical and ethical challenges; 4) five‑year outlook. Use clear headings and transition sentences."

2. Provide Output Samples

Appending 1‑2 complete examples improves format accuracy by 45%.

Original: "Generate a weather reply."\nStructured: "When a user asks for weather, reply in the following format:\n【City】: Beijing\n【Date】: 2026‑02‑03\n【Temperature】: -2°C to 5°C\n【Condition】: Partly cloudy → sunny\n【AQI】: Good (75)\n【Advice】: Wear layers in the morning; air quality is good for outdoor activities."

3. Set Constraints

Explicit limits cut hallucination rates by 60%.

# Role\nYou are a health‑information advisor, not a medical professional.\n# Constraints\n- Do not provide diagnoses, prescriptions or treatment advice.\n- All information must be based on provided authoritative sources.\n- If a professional answer is required, reply "Please consult a professional doctor."

4. Design a Thought‑Chain Workflow

A clear step‑by‑step reasoning chain can increase complex‑problem accuracy by >50%.

Original: "Plan a trip for me."\nStructured: \n1. Confirm basic info (departure, destination, dates, participants, budget).\n2. Query destination weather via plugin.\n3. Recommend attractions based on weather.\n4. Design itinerary, allocate time, suggest meals and rest.\n5. Handle exceptions (fallback weather, generic spots, missing info).

5. Optimize Format and Structure

Using Markdown or XML improves readability and parsing efficiency by ~30%.

<response>\n <issue_type>{{type}}</issue_type>\n <product_model>{{model}}</product_model>\n <solution>\n 1. {{step1}}\n 2. {{step2}}\n 3. {{step3}}\n </solution>\n <prevention>{{advice}}</prevention>\n</response>

Building and Managing a Personal Prompt Library

Classification System (Knowledge Graph)

Organize prompts by function, scenario, domain and complexity. Studies show a good classification boosts retrieval efficiency by 75%.

Template‑Based Storage

Extract common parts into templates and keep variable parts as parameters (e.g., {{input}}, {{model}}). This "design once, use many times" approach can shorten new‑project development time by 60%.

Version Control and Iterative Optimization

Maintain independent version history for each prompt, record modification reasons and test results, and run A/B tests. Version control improves iteration efficiency by 50%.

v1.0.0 (2025‑01‑20) – Initial version, simple role, no workflow.\nTest: accuracy 78%, format correctness 65%\n\nv1.1.0 (2025‑02‑05) – Added skills and workflow.\nTest: accuracy 89%, format correctness 92%\n\nv1.2.0 (2025‑03‑15) – Introduced variables and conditional logic.\nTest: accuracy 93%, format correctness 98%

Team Collaboration Mechanism

Use shared spaces, role‑based access control (Admin, Developer, Tester, Guest) and a prompt‑review workflow. Effective collaboration can raise overall team efficiency by 30%.

Automation Tools Support

Coze Studio’s Prompt Management API enables automated updates, version comparison, testing and monitoring alerts. Automation can increase management efficiency by 40%.

Daily automatic backup of the prompt library to cloud storage.

Weekly auto‑generated usage reports.

Automatic A/B testing when a new version is released.

Trigger optimization workflow when user satisfaction falls below a threshold.

Periodic auto‑generation of prompt‑library documentation.

Practical Case: Enterprise‑Level IT‑Policy Q&A Agent

Requirement Analysis and Framework Selection

XX Technology needs an agent that accurately answers employee questions about IT policies (device procurement, network security, software licensing, etc.). The task requires multi‑domain handling, strict adherence to internal documents, clarification of ambiguous queries, and a user‑friendly tone for non‑technical staff. The R‑C‑S‑W‑O model is chosen for its comprehensive coverage.

Prompt Design and Optimization

Initial Prompt :

# Role\nYou are the AI assistant of the IT department at XX Technology.\n# Background\nAnswer employee questions based on the provided IT‑policy documents.\n# Skills\n1. Call the IT‑policy knowledge base.\n2. Parse keywords from the user query.\n3. Identify the policy category.\n# Workflow\n1. Receive input.\n2. Identify question type.\n3. Query knowledge base.\n4. Generate answer.\n# Constraints\n- Only answer IT‑policy questions.\n- Base answers on provided documents.\n- No personal opinions or speculation.\n- Use professional yet easy‑to‑understand language.

Optimization Steps :

Structure background into three clearly labeled document fragments.

Detail skill triggers and execution methods.

Expand workflow with multi‑step handling and exception mechanisms.

Define output using XML tags for machine‑readable structure.

Add version‑control entries for each modification.

Final Optimized Prompt (excerpt) :

# Role\nYou are the AI assistant of XX Technology’s IT department, responsible for answering questions about the "IT Equipment Management and Information Security" manual.\n\n# Background\nPlease answer based on the following three document fragments:\n---Fragment 1---\nContent: [retrieved text 1]\n---Fragment 2---\nContent: [retrieved text 2]\n---Fragment 3---\nContent: [retrieved text 3]\n\n# Skills\n1. Retrieve relevant information from the knowledge base.\n2. Analyze keywords and intent.\n3. Identify the policy category (device procurement, network security, software licensing, etc.).\n\n# Workflow\n1. Analyze the question.\n2. Locate relevant information.\n3. Judge answer completeness:\n - If sufficient, summarize and answer.\n - If partially relevant, answer with noted limitations.\n - If unrelated, reply "I cannot answer based on the IT‑policy documents."\n4. Generate answer.\n5. Handle exceptions (unrecognizable type, knowledge‑base failure, missing documents).\n\n# Output\n <response>\n <answer>{{concise answer}}</answer>\n <detail>{{additional context}}</detail>\n <source>{{policy section reference}}</source>\n <format>{{professional business language}}</format>\n <length>{{≤150 characters}}</length>\n</response>

Testing and Iterative Optimization

Test scenarios:

Basic functionality – questions directly covered by the documents.

Ambiguous queries – fuzzy or incomplete questions.

Cross‑domain queries – questions outside IT‑policy scope.

Multi‑turn dialogue – consecutive related questions.

Exception handling – simulate knowledge‑base unavailability.

A/B Test Results :

v1.0.0: Accuracy 78%, format correctness 65%, user satisfaction 3.2/5.

v1.1.0: Accuracy 89%, format correctness 92%, user satisfaction 3.8/5.

v1.2.0: Accuracy 93%, format correctness 98%, user satisfaction 4.3/5.

Analysis shows that adding structured background, detailed skills and XML output significantly improves both accuracy and user satisfaction.

Summary and Outlook

Structured prompt design is a core competency for AI agent development. By providing clear logic, strict constraints and explicit output specifications, structured prompts enable large language models to execute tasks more precisely and deliver higher‑quality responses.

Framework suitability:

R‑C‑S‑W‑O – best for complex, enterprise‑level tasks.

Zeng Yingjie three‑stage – ideal for rapid prototyping and beginners.

CRISPE – excels when fine‑grained output style control is required.

Applying the five optimization techniques (clarify context, provide examples, set constraints, design thought‑chains, optimize format) can markedly improve prompt effectiveness and agent performance.

Future direction: prompt design will evolve toward greater modularity, automation and integration with platform tools such as Coze’s upcoming AgentPlan feature (expected Jan 2026), which will enable agents to decompose long‑term goals, devise strategies and execute them cyclically.