📝 Writing Effective Prompts

To create AI Applications that consistently produce quality content, you will need to write effective prompts.

AI is a tool, not black magic. It must be guided through precise instructions and fed with relevant information to write satisfactory responses.

Key elements

There are a multitude of prompting techniques with advantages and disadvantages, which will be more or less adapted to your needs.

We can nevertheless identify common points across different methods:

🎭 Role and context

The importance of role and context: The more you give AI a reference framework, the better it will be able to understand what is expected of it, the positioning it must adopt, and the interpretation it will have of your directives.

🎯 Precision of terms

Precision of terms used: Be factual, precise, avoid words that can be subject to interpretation. For example, for AI to write a "good" text means nothing without criteria to determine what you consider "good", asking for "many" words is also not quantifiable.

📚 Use of references

Use of references: Give it material it can rely on, examples of expected content by explaining what the provided excerpt represents.

🗂️ Use of structure

Use of structure: See the section below on this page, AI is much more relevant when we explicitly distinguish each element it must take into account and an order is established for the logic to follow. You can even go so far as to explicitly indicate the different thinking steps it must follow before writing its content.

⚠️ Optimal prompt length

A prompt that is too short will lack information to be properly interpreted by AI, and a prompt that is too long risks confusing it with too much information to understand. To avoid lengthy prompts, break them down into subtasks.

🎯 Single-purpose design

Take advantage of Thinkeo and AI blocks. AI is most effective when we break down what it needs to do into several subtasks, so consider that a Prompt is intended for a single objective.

🔗 Thinkeo block chaining

Take advantage of Thinkeo's features to create cascading reasoning. Use interconnected AI blocks to break down complex tasks into simpler subtasks. Share blocks to reuse the same references, contexts, or prompt sections from one prompt to another. Nest AI blocks within others to leverage content generated by your other AI blocks, thus unifying the result of your subtasks to accomplish a larger one.

Additional best practices

Testing & examples

Use concrete examples: Provide specific examples of the type of response you expect. This helps AI better understand your expectations and calibrate its response accordingly.

Test your prompts: Before finalizing a prompt for an application, test it with different input variations to ensure it produces consistent and quality results in various situations.

Iterate and refine: After testing your prompt, don't hesitate to adjust it. Sometimes, small changes in wording can have a significant impact on response quality.

Advanced techniques

Use the "AI as expert" technique: Ask AI to adopt the viewpoint of a specific expert in the relevant field. For example, "As a digital marketing expert with 20 years of experience, how would you approach...". This approach can help obtain more nuanced and professional responses.

Incorporate creative constraints: Sometimes, adding unusual constraints can stimulate more original and insightful responses. For example, "Explain this concept as if you were talking to a 10-year-old" or "Describe this strategy using only culinary metaphors". These constraints can lead to clearer explanations or unique perspectives.

Thinkeo Pro Tip

💡 Use shared blocks in Thinkeo to avoid repeating the same information in your prompts. A Choice Block for Tone can be nested in all your AI blocks, conditional blocks with certain other instructions as well, use blocks for content given as "provided information", etc.

Consult the AI Block page to see an illustration.

AI block architecture in Thinkeo

Understanding how AI blocks work in Thinkeo is essential for creating effective prompts. In Thinkeo, all nested blocks within an AI block constitute the complete prompt that will be sent to the AI.

How prompt assembly works

📝 Block concatenation

When an AI block executes, Thinkeo automatically assembles all its nested blocks from top to bottom to create the final prompt. Each nested block adds its content to the complete prompt.

📁 Automatic content injection

File blocks automatically extract and inject document content into the prompt - no special syntax needed. The AI receives the extracted text as part of the assembled prompt.

🔄 Child block execution

All nested blocks (File, Search, Paragraph, etc.) execute first and generate their content before the parent AI block receives the complete assembled prompt.

Practical example

AI Block: "Document Analysis"
├── 📝 Paragraph Block: "You are a financial expert. Analyze the document below:"
├── 📁 File Block: [Automatically extracts: "Q3 Revenue Report... $2.3M increase..."]
└── 📝 Paragraph Block: "Provide 3 key insights in numbered format."

The AI receives this complete prompt:

You are a financial expert. Analyze the document below:

Q3 Revenue Report... $2.3M increase... [full extracted content]

Provide 3 key insights in numbered format.

Nested AI block workflows

You can create sophisticated workflows by nesting AI blocks within other AI blocks:

🔍 Child AI generates content

The nested AI block executes first, generates its response, and this response becomes visible content for the parent AI block.

📊 Parent AI uses child results

The parent AI block can then use, synthesize, or build upon the content generated by its child AI blocks.

Example Workflow:

Parent AI Block: "Executive Summary"
├── 📝 Instructions: "Create an executive summary based on the analysis below:"
├── 🤖 Child AI Block: "Financial Analysis" 
│   ├── 📝 Instructions: "Analyze financial data and identify 3 key trends"
│   └── 📁 File Block: [Financial reports]
└── 📝 Final Instructions: "Format as 200-word executive summary"

Variables and dynamic content

Text variables in Thinkeo

Only text attributes can be used as variables within paragraph blocks using the @@ syntax.

🎯 Variable insertion

Type @@ to open a search window and select your desired text attribute. Once inserted, the attribute appears as a colored button-like element in your block showing the attribute name.

📝 Visual representation

In your blocks, variables appear as: [industry_sector] (colored button-style elements), not as @@industry_sector@@. The @@ syntax is only used during insertion.

✅ Runtime replacement

During block execution, these colored attribute buttons are replaced with their actual values assigned during the user's publication journey.

File attributes usage

File attributes cannot be used directly with @@ syntax. Instead:

File attributes are assigned to File Blocks or Search Blocks in their parameters

These blocks automatically extract and process file content

The processed content becomes part of the prompt assembly

Best practices for variables

📋 Descriptive naming

Use clear, descriptive names for your attributes: [client_company], [project_deadline], [industry_sector] (displayed as colored buttons)

🔄 Fallback instructions

Include instructions for empty variables: "If company name is not provided, use 'the organization'"

🎯 Strategic placement

Place variables strategically in your prompt structure for maximum impact and clarity

🧪 Test different values

Use the Studio's Attributes tab to test your prompts with different variable values

Structure

An effective prompt needs structure to organize information, provide context, and guide AI understanding for each instruction. It's not necessary to always have the same one; however, having a basic foundation that you use recurrently will help you build your prompts and adapt them to each need.

Recommended basic structure

We recommend a basic structure that has proven itself on Thinkeo Apps and that we adapt to uses. You can use it, be inspired by it, or start with something completely different - this is a proposal to help you get started.

🎭 Role

Role: AI's role, we essentially indicate the Position it occupies and in what context or company. This role is generally common across the entire App.

Example: "You are a B2B marketing expert specializing in the [industry_sector] industry with 15 years of experience." (where [industry_sector] appears as a colored button)

🎯 Objective

Objective: Detailed instruction that will specify to AI what we expect from it. Here we will specify the context if needed, indicate how to interpret the content we provide, explain what it must produce and the objective of what will be written.

Example: "Analyze the provided document to create a competitive analysis of 300 words that identifies 3 key differentiating advantages and 2 potential market opportunities."

📄 Provided information

Provided information: The context we will give to AI and the references it can rely on. This can be words, sentences, or paragraphs. This is where we can share content to feed several prompts in parallel, or feed an AI block with content generated by another, notably.

Example: "Variables used: Company: [client_company], Competitor: [main_competitor]. Document to analyze:" (variables appear as colored buttons)

🎨 Tone

Tone: Structured in three distinct points: Language level / Editorial style / Tone

Example: "Professional and analytical tone. Use business terminology. Be precise and factual."

📏 Length

Length: The number of words or sentences not to exceed and the expected structure, like X sentences or X paragraphs.

Example: "300 words maximum, structured in 3 numbered sections with bold headings."

🗣️ Language

Language: Here we specify specific rules like mandatory formal address, inclusive writing, particular words to avoid or reformulate.

Example: "Use formal address throughout. Avoid jargon. Replace technical terms with accessible explanations."

🚫 Prohibitions

Prohibitions: AI handles prohibitions poorly, it's much more effective to give precise instructions on all other points; here it's about adding a complementary directive on specific elements.

Example: "Do not make recommendations about pricing or budget allocation."

💡 Bonus tip: The instruction "Let's think step by step." can help when placed between "Role" and "Objective", AI then takes each instruction as a step, which often allows obtaining more qualitative results.

Alternative structures

Here are other proven structures you can use or adapt according to your needs:

Classic alternative structure

Role

Context

Objective (mandatory)

Rules

Language (tone is part of language)

Length

Prohibitions

Provided information

STAR method structure

Particularly effective for analysis tasks:

Situation - Context and background

Task - What needs to be accomplished

Action - How to approach the task

Result - Expected output format

CRISP structure

For structured analytical work:

Context - Business/domain context

Requirements - Specific needs and constraints

Input - Data and information provided

Scope - Boundaries and limitations

Process - Step-by-step methodology

Complex task structure

For detailed analysis requiring high precision:

Role

Context

Document structure

Instructions

Objective

Content to include

Tone

Length

Structure

10.

Specific rules

11.

Provided information

Rapid response structure

For simple, fast tasks:

Role - Brief expert definition

Task - One clear objective

Format - Output specification

Data - Input information

Iterative refinement structure

For multi-step improvement tasks:

Initial role - Starting position

Analysis phase - First assessment

Refinement criteria - Improvement guidelines

Final output - Polished result format

Practical use cases and examples

Marketing content creation

Scenario: Creating personalized marketing content based on client industry

AI Block Structure:
├── 📝 "You are a marketing expert specializing in [client_industry]"
├── 📝 "Create a compelling value proposition for [client_company]"
├── 📁 File Block (client brief document)
└── 📝 "Format as 3 bullet points with impact metrics"

(Variables appear as colored buttons: [client_industry], [client_company])

Document analysis workflow

Scenario: Multi-stage document analysis with specialized AI agents

Parent AI Block: "Executive Summary"
├── 📝 "Synthesize the following analyses into an executive summary:"
├── 🤖 Child AI: "Financial Analysis"
│   ├── 📝 "Analyze financial data and identify trends"
│   └── 📁 Financial reports
├── 🤖 Child AI: "Risk Assessment" 
│   ├── 📝 "Identify potential risks and mitigation strategies"
│   └── 📁 Same financial reports
└── 📝 "Present findings in 250 words with clear recommendations"

Prompt quality checklist

Before finalizing any prompt, verify:

Testing and debugging your prompts

Using Studio's debug features

🏷️ Attributes tab testing

Navigate to the Attributes tab in Studio to assign test values to your text attributes. This allows you to simulate different scenarios without creating actual publications.

▶️ Individual block testing

Use the "play" button next to block parameters to execute individual blocks and verify their output before testing the complete workflow.

👁️ Output visualization

Toggle "show output" to see execution results directly in Studio, helping you identify issues and verify expected behavior.

🔄 Iterative refinement

Test one modification at a time, compare results, and refine systematically rather than making multiple changes simultaneously.

Common issues and solutions

🔴 Variables not replacing

Symptoms: Variable buttons still show attribute names instead of values
Solution: Verify attribute exists, check spelling, ensure it's assigned in wizard step

🔴 Inconsistent responses

Symptoms: AI generates different styles/formats
Solution: Add more specific formatting instructions, include examples, adjust temperature parameter

🔴 Missing file content

Symptoms: AI doesn't reference document content
Solution: Check file block configuration, verify file attribute assignment, test with different file formats

🔴 Prompt too long

Symptoms: Truncated responses, high costs
Solution: Break into smaller AI blocks, use child/parent structure, choose model with larger context

Diagnostic questions

When a prompt isn't working effectively, ask yourself:

Is the role specific enough? Generic roles produce generic results

Are instructions actionable? Avoid vague terms like "good" or "professional"

Is the block hierarchy logical? Check execution order and dependencies

Are length constraints realistic? Match expectations with model capabilities

Have I tested edge cases? Try with different variable values and file types

Choosing the right model for your prompt

Model selection strategy

The complexity and requirements of your prompt should guide your model choice:

📝 Simple content generation

Use case: Basic content creation, simple analysis, straightforward formatting

Recommended models: GPT-4.1 Mini, Claude 3.5 Haiku, Mistral Small
Why: Cost-effective for simple tasks, fast execution, sufficient quality

🧠 Complex analysis & reasoning

Use case: Multi-step analysis, complex reasoning, sophisticated synthesis

Recommended models: GPT-4.1, Claude Sonnet 4, o3 Mini (with reasoning)
Why: Better reasoning capabilities, handle complex instructions, maintain consistency

📊 Large document processing

Use case: Processing very large documents, extensive context analysis

Recommended models: Gemini 2.5 Pro (2M tokens), GPT-4.1 (1M tokens)
Why: Massive context windows, efficient large document handling

🔍 Research & web integration

Use case: Tasks requiring current information, fact-checking, web research

Recommended models: Perplexity Sonar Pro, Perplexity Deep Research
Why: Integrated web search capabilities, up-to-date information access

Cost optimization guidelines

💡 Efficiency strategies

Start with smaller models for testing and refinement

Use GPT-4.1 for best quality/price ratio on complex tasks

Reserve premium models (Gemini 2.5 Pro) for truly large documents

Consider open-source options for less critical applications

📈 Performance indicators

If responses are inconsistent → Try more powerful model

If costs are too high → Test with smaller model or break into subtasks

If processing is too slow → Consider faster alternatives

If context is truncated → Upgrade to larger context model

A/B testing models

To find the optimal model for your use case:

Define success metrics: Quality, consistency, speed, cost

Test with representative data: Use typical inputs and scenarios

Compare results: Document differences in output quality and characteristics

Calculate ROI: Balance quality improvements against cost increases

Make informed decision: Choose based on your specific requirements and budget

Mastering temperature for better results

Temperature is one of the most powerful and underutilized parameters in prompting. Understanding its impact can dramatically improve your results.

Understanding temperature

Temperature controls the "creativity" vs "predictability" balance of AI responses:

❄️ Low temperature (0.3-0.5)

Consistent & professional

Reliable for business content and analysis

Natural language without being robotic

Good balance of consistency and readability

Recommended baseline for most Thinkeo Apps

🔥 Higher temperature (0.6-0.9)

Creative & varied

More diverse and creative outputs

Better for brainstorming, marketing copy, storytelling

Each generation will be more unique

Use when you specifically need originality

Practical temperature guidelines

📊 For analysis & data (temperature: 0.3-0.4)

When processing documents, extracting information, or creating structured analyses. You want consistency and accuracy while maintaining some natural variation.

Example: "Extract 5 key financial metrics from this quarterly report"

✍️ For content creation (temperature: 0.4-0.6)

For writing marketing content, proposals, or explanations where you want balanced creativity with maintained coherence. This is the sweet spot for most business content.

Example: "Write a compelling value proposition for [product_name]"

💡 For brainstorming (temperature: 0.6-0.8)

When you want innovative ideas, creative solutions, or diverse perspectives on a problem while maintaining practicality.

Example: "Generate 10 innovative marketing campaign ideas for [industry_sector]"

🎨 For creative writing (temperature: 0.7-1.0)

For storytelling, creative descriptions, or when you want maximum originality and unique perspectives.

Example: "Write a creative product description that stands out from competitors"

Temperature best practices

Pro tips:

Start with 0.3-0.4 and adjust: Begin with a moderate temperature and increase for more creativity

Test systematically: Try the same prompt with different temperatures to see the impact

Match temperature to task: Analysis = 0.3-0.4, business content = 0.4-0.6, creativity = 0.6+

Use App-level settings: Set a baseline temperature for your entire App, then adjust individual blocks as needed

Temperature testing example

Try this same prompt with different temperatures to see the difference:

Prompt: "Describe the benefits of remote work for companies"

Temperature 0.3: "Remote work reduces office costs, increases talent pool access, and improves employee satisfaction through flexible scheduling..."

Temperature 0.7: "Picture this: your company becomes a borderless ecosystem where talent flourishes from kitchen tables to mountain cabins, cutting costs while boosting creativity..."

The same prompt, completely different tone and approach!

Going further

The Prompt Engineering guide proposed by OpenAI is a good foundation for discovering different approaches and best practices to implement in Prompts addressed to OpenAI models.

📚 External resource:
OpenAI Platform
Claude

Keep in mind that iterating and adjusting your prompts based on your usage remains the best practice for understanding and seeing what works best for your needs. The recommendations you'll find for OpenAI models also apply to other AIs (like Mistral for example), so once you've found your bearings and designed a standard structure that brings you satisfaction, reuse it.

Key elements#

Additional best practices#

Testing & examples#

Advanced techniques#

AI block architecture in Thinkeo#

How prompt assembly works#

Practical example#

Nested AI block workflows#

Variables and dynamic content#

Text variables in Thinkeo#

File attributes usage#

Best practices for variables#

Structure#

Recommended basic structure#

Alternative structures#

Classic alternative structure#

STAR method structure#

CRISP structure#

Complex task structure#

Rapid response structure#

Iterative refinement structure#

Practical use cases and examples#

Marketing content creation#

Document analysis workflow#

Prompt quality checklist#

Testing and debugging your prompts#

Using Studio's debug features#

Common issues and solutions#

Diagnostic questions#

Choosing the right model for your prompt#

Model selection strategy#

Cost optimization guidelines#

A/B testing models#

Mastering temperature for better results#

Understanding temperature#

Practical temperature guidelines#

Temperature best practices#

Temperature testing example#

Going further#

Key elements

Additional best practices

Testing & examples

Advanced techniques

AI block architecture in Thinkeo

How prompt assembly works

Practical example

Nested AI block workflows

Variables and dynamic content

Text variables in Thinkeo

File attributes usage

Best practices for variables

Structure

Recommended basic structure

Alternative structures

Classic alternative structure

STAR method structure

CRISP structure

Complex task structure

Rapid response structure

Iterative refinement structure

Practical use cases and examples

Marketing content creation

Document analysis workflow

Prompt quality checklist

Testing and debugging your prompts

Using Studio's debug features

Common issues and solutions

Diagnostic questions

Choosing the right model for your prompt

Model selection strategy

Cost optimization guidelines

A/B testing models

Mastering temperature for better results

Understanding temperature

Practical temperature guidelines

Temperature best practices

Temperature testing example

Going further