The Problem I Was Trying to Solve§
As a developer juggling multiple projects, I found myself repeatedly writing the same kinds of prompts for different LLMs. Whether it was generating boilerplate code, refactoring messy functions, or writing documentation, I'd spend 10–15 minutes crafting the perfect prompt each time. Worse, I'd often forget the exact phrasing that worked best for a specific tool like DeepSeek versus Claude. This inefficiency was eating into my actual coding time.
I realized that what I needed was a set of "cheat codes"—pre-defined, battle-tested prompts that I could copy-paste into any LLM interface and get consistent, high-quality results. The goal was to reduce prompt engineering overhead to near zero, allowing me to focus on the logic and architecture of my code.
Tools and Setup§
I rely on a mix of LLM tools: DeepSeek for its strong coding abilities and low cost, Claude for complex reasoning tasks, and Cursor as my integrated development environment with built-in AI. For quick research or debugging, I use Perplexity. Each tool has its own strengths, and I wanted prompts that played to those.
To organize my prompts, I created a simple markdown file (prompt_vault.md) with sections for each task type. I use VSCode with a plugin that allows quick snippet insertion, but even a plain text file works. The key is having the prompts readily available for copy-paste.
Step-by-Step: What I Actually Did§
Step 1: Identify High-Impact Tasks§
I listed the most common coding tasks I offload to LLMs: writing unit tests, generating SQL queries, refactoring functions, creating API endpoints, and explaining legacy code. For each, I defined the exact output I wanted.
Step 2: Craft Prompts with Tool-Specific Tweaks§
For each task, I wrote a base prompt and then adapted it for each tool. For example, for DeepSeek, I add "Think step by step and output concise code" because DeepSeek tends to be verbose. For Claude, I use "Provide a thorough explanation with code" since Claude excels at reasoning.
Step 3: Test and Iterate§
I ran each prompt on real codebases and adjusted based on output quality. I kept what worked and discarded the rest. This iterative process took about two weeks but resulted in a vault of 50+ high-quality prompts.
Step 4: Organize for Copy-Paste§
I grouped prompts by category and added clear titles. I use a consistent format: first line is the task description, then the system message (if any), then the user message placeholder. For example:
## Generate Python Unit Test for Function
**System**: You are an expert Python developer specialized in unit testing using pytest. Follow best practices and use mocks where appropriate.
**User**: Write a comprehensive unit test for the following function. Include edge cases: `{function_code}`This structure lets me quickly copy and paste the specific blocks into any LLM chat.
Code Samples / Prompts Used§
Here are a few of my most-used cheat codes:
Prompt 1: Refactor Function (Works best with DeepSeek)
Refactor this Python function to be more performant and readable. Use type hints, docstrings, and break it into smaller helper functions if needed. Output only the final code.
{function_code}
Prompt 2: Explain Legacy Code (Works best with Claude)
I'm a junior developer. Explain what this code does step by step. Use analogies to make it understandable. Then suggest improvements without changing the behavior.
{code_block}
**Prompt 3: Generate SQL Query (Works best with Perplexity)**
Given the following table schema:
{create_table_statements}
:::sandbox
prompt_id: 2bde446f-be4a-c847-2f9c-650103ac583c
tool_name: Perplexity AI
label: Try this Perplexity AI prompt →
:::
Write an SQL query that {describe_question}. Optimize for performance. Explain your choices.**Prompt 4: Create API Endpoint (Works best with Cursor's inline AI)**
In this FastAPI project, create a new endpoint to {description}. Use Pydantic models for request/response, add validation, and include error handling. Follow the existing code style.What Worked Well§
First, the time savings were dramatic. What used to take 15 minutes of prompt engineering now takes 10 seconds of copy-paste. My prompt vault reduced the cognitive load of switching between tools because each prompt is pre-tuned for the target LLM.
Second, consistency improved. Because I'm using the same prompt structure, the output style is more uniform. This is especially helpful when generating documentation or unit tests across different parts of the codebase.
Third, the "cheat codes" serve as a personal knowledge base. When I revisit a project after months, I can quickly get back up to speed by using my explanation prompts.
What Failed and Why§
Not everything worked. Prompts that were too generic (e.g., "Write code for this problem") produced mediocre results. The fix was to be extremely specific about the output format, constraints, and context.
Another failure was trying to use the same prompt for different LLMs without adjustments. For example, a prompt that works well with Claude might cause DeepSeek to hallucinate or produce overly verbose code. I had to create tool-specific variants.
Finally, I initially stored prompts in a complex Notion database with tags and metadata. That added friction—I spent more time navigating the database than actually using the prompts. Moving to a simple flat markdown file solved that.
Results and Takeaways§
After three months, I've saved approximately 40 hours of prompt engineering time. My code quality has improved because I consistently apply best practices via the prompts. I now have a reusable asset that I can update as new LLM features roll out.
Key Takeaways:
- Curate, don't just collect: Only keep prompts that have been tested and work reliably.
- Optimize per tool: Each LLM has quirks—tailor prompts accordingly.
- Keep it simple: A plain text file with clear sections is more effective than a complex system.
- Iterate continuously: As models update, revisit and refine your prompts.
Try It Yourself§
Start by listing your top 5 repetitive coding tasks. Write one prompt for each, test it on your preferred LLM, and tweak until you get consistent results. Over a week, you'll have a mini vault that pays dividends.
Share your own cheat codes on LLMDB.APP and help the community build efficiency. Remember, the goal is to spend less time talking to AI and more time building with it.
