The capabilities of modern Large Language Models (LLMs) have evolved from simple query engines into sophisticated reasoning partners. However, the quality of the output remains strictly bound by the quality of the input. For professional prompt engineers, the goal is no longer just writing instructions; it is about architecting systems that force reliability, reasoning, and adherence to constraints.
The following prompts have been rigorously tested and optimized for ChatGPT, Gemini, Claude, and DeepSeek. While each model possesses distinct architectural advantages—such as DeepSeek’s logical rigor or Claude’s semantic nuance—these 10 scripts provide a universal foundation for high-level prompt engineering. They leverage the AI itself to refine, debug, and optimize your interactions.
1. The Recursive Refiner
Best for: Claude for its ability to adhere to safety guidelines and nuanced instruction.
This prompt turns the AI into a critical editor, analyzing your draft prompt for ambiguities and suggesting specific improvements before execution.
Act as a Senior Prompt Engineer. I am going to provide you with a draft prompt I intend to use. Your goal is to critique it based on the following criteria:
1. Clarity and Ambiguity
2. Constraint Adherence potential
3. Edge Case vulnerabilities
After the critique, rewrite the prompt into a highly optimized version using best practices (e.g., chain-of-thought, delimited context).
Draft Prompt: "[INSERT YOUR DRAFT PROMPT HERE]"
The Payoff: This eliminates the trial-and-error phase by identifying logical loopholes before you ever run the prompt in production.
2. The Chain-of-Thought Architect
Best for: DeepSeek for its strength in handling complex logic and mathematical reasoning.
Zero-shot prompts often fail at complex reasoning. This prompt generates the specific step-by-step logic the AI needs to follow to solve a problem accurately.
I need to solve the following complex task: [INSERT TASK DESCRIPTION].
Create a "Chain of Thought" (CoT) framework for this task. Break down the reasoning process into discrete, logical steps that an AI model should internalize before generating the final answer. Output the result as a system instruction that starts with "Let's think step by step..."
The Payoff: By explicitly defining the reasoning path, you significantly reduce hallucinations and logic errors in complex workflows.
3. The Few-Shot Generator
Best for: Gemini for its large context window and pattern recognition capabilities.
Providing examples (few-shot prompting) is the most effective way to guide style and format. This prompt automates the creation of high-quality training examples.
I need 5 distinct pairs of "Input" and "Ideal Output" examples to teach an AI model how to perform the following task: [INSERT TASK].
The examples must cover:
1. A standard use case.
2. An edge case (missing or messy data).
3. A complex use case requiring inference.
Format these as a JSON object suitable for few-shot context insertion.
The Payoff: Instantly creates the data necessary to “train” the model within the context window, ensuring consistent formatting and tone.
4. The Persona Simulator
Best for: ChatGPT for its versatility in creative writing and role adoption.
A generic “Act as an expert” instruction is often too vague. This prompt builds a deep, psychological profile for the AI to inhabit.
Define a comprehensive system persona for an AI agent tasked with [INSERT JOB/ROLE].
Include:
1. Tone and Voice guidelines (adjectives only).
2. Explicit "Do Not" constraints.
3. A specific domain knowledge base it should prioritize.
4. A behavioral directive for how to handle uncertainty (e.g., ask clarifying questions vs. make assumptions).
Output this as a strictly formatted System Message block.
The Payoff: Transforms the AI from a generalist into a specialist agent with predictable behaviors and boundary conditions.
5. The Constraint Stress Test (Red Teaming)
Best for: DeepSeek or Claude for their robust adherence to negative constraints.
Before deploying a prompt, you must know how it fails. This prompt attempts to “break” your instructions to find weaknesses.
Here is a prompt I have designed: "[INSERT PROMPT]".
Act as a "Red Team" adversary. Your goal is to generate 5 specific user inputs that would likely cause this prompt to fail, produce hallucinations, or bypass the intended constraints. Explain why each input is a vulnerability.
The Payoff: Proactively identifies security risks and logic gaps, allowing you to patch them before end-users interact with the system.
6. The Token Compressor
Best for: ChatGPT or DeepSeek for efficient phrasing.
When working with API costs or limited context windows, verbosity is expensive. This prompt retains the instruction’s intent while minimizing token count.
Below is a verbose system prompt. Rewrite it to be as concise as possible without losing any instructional nuance or constraints. Focus on high-density language and imperative verbs.
Original Prompt: "[INSERT VERBOSE PROMPT]"
The Payoff: Reduces operational costs and latency while keeping the context window open for more important user data.
7. The Socratic Requirement Gatherer
Best for: Claude for its conversational depth and ability to guide user intent.
Use this when you (or your client) define a vague goal. It forces the AI to ask you questions to clarify the objective before writing the prompt.
I want to build a prompt to achieve [INSERT VAGUE GOAL], but I am unsure of the specific requirements.
Do not generate the prompt yet. Instead, ask me 5 targeted questions that will help clarify the audience, format, constraints, and success metrics. Once I answer, you will use that information to construct the perfect prompt.
The Payoff: Prevents the “Garbage In, Garbage Out” cycle by ensuring the specifications are clear before any prompt engineering begins.
8. The Cross-Model Translator
Best for: ChatGPT for its broad understanding of other model behaviors.
A prompt that works in one model may fail in another. This prompt adjusts syntax and structure for specific model architectures.
I have a prompt optimized for [Model A]. Please rewrite this prompt to be optimized for [Target Model: e.g., DeepSeek/Claude].
Adjust for [Target Model]'s specific preferences regarding:
- XML tag usage vs. Markdown.
- System prompt placement.
- Verbosity and refusal triggers.
Original Prompt: "[INSERT PROMPT]"
The Payoff: Ensures your prompt library remains agnostic and effective regardless of the underlying infrastructure changes.
9. The Output Validator (The Judge)
Best for: Gemini or Claude for high-level semantic analysis.
In automated workflows, you need a second AI to score the output of the first. This prompt creates a rubric for automated evaluation.
Create an evaluation rubric to score AI responses for the task: [INSERT TASK].
The output should be a strictly formatted JSON object containing:
1. A score from 1-10.
2. A binary 'Pass/Fail' flag based on critical constraints.
3. A brief reason for the score.
The criteria for success are: [INSERT CRITERIA].
The Payoff: enables scalable, automated quality assurance for AI applications without requiring human review for every interaction.
10. The Dynamic Variable Template
Best for: DeepSeek for structured coding tasks.
Hard-coded prompts are not scalable. This prompt converts a static text into a template with clear variables for programmatic insertion.
Analyze the following text interaction and convert the prompt into a reusable template.
Replace all specific data points (names, dates, specific topics) with distinct variable placeholders using the {{VARIABLE_NAME}} format. List the variables at the end with a description of what data type should populate them.
Input Prompt: "[INSERT STATIC PROMPT]"
The Payoff: Facilitates the integration of prompts into software applications, allowing for dynamic content injection via API.
Pro-Tip: Context Delimiters
When constructing complex prompts, always separate instructions from data using clear XML-style delimiters (e.g., <context>, <instructions>, <user_input>). This prevents “prompt injection” confusion where the model mistakes user data for system instructions. Claude and DeepSeek are particularly responsive to this structure, resulting in significantly higher accuracy.
Mastering prompt engineering is not about memorizing magic phrases; it is about understanding how different architectures process information. By utilizing these meta-prompts, you shift from manually writing instructions to engineering robust, scalable interaction frameworks. Test these across ChatGPT, Gemini, Claude, and DeepSeek to build a resilient toolkit that adapts to any challenge.