In this guide you will find useful information about setting up APIs for different LLMs and also how to prompt them effectively.
- Prompt Engineering
Prompt engineering is a new and emerging art of crafting instructions to get LLMs like ChatGPT, Claude, Gemini and others, to produce the output you want. The better the instructions, the better the results!
Review this [LANGUAGE] code for:
- Bugs and logical errors
- Performance issues
- Security vulnerabilities
- Code style and best practices
Code:
```[LANGUAGE]
[CODE]
```
Provide specific line-by-line feedback.You are a startup CTO.
My idea:
<DESCRIBE YOUR IDEA>
I have 48 hours.
Generate:
1. Feature set
2. System architecture
3. Folder structure
4. First 5 implementation steps
Keep everything minimal.You are an expert QA engineer and software tester.
Given this function or module:
<PASTE CODE>
Do the following:
1. Identify all normal cases and edge cases.
2. Generate unit tests covering these cases.
3. Use the language and testing framework specified: <LANGUAGE + FRAMEWORK>.
4. Ensure tests are ready to run (copy-pasteable).
5. Label each test clearly.
6. If the code has potential bugs or risky behavior, include a failing test case to catch it.You are a senior frontend designer and UX engineer.
We are building:
<PROJECT IDEA>
Do the following:
1. Generate a list of screens/components needed.
2. Suggest a simple component hierarchy.
3. Recommend Tailwind (or CSS) classes for styling.
4. Provide a minimal mobile-first layout plan.
5. Include placeholder content for testing.You are an expert software engineer.
Given this code:
<PASTE CODE>
Do the following:
1. Refactor for readability and performance.
2. Add error handling where appropriate.
3. Remove any dead or redundant code.
4. Keep the output minimal and functional.
Return only the improved code.Vague prompts get vague results. Include details such as the language, format, length, style and constraints.
❌ Weak:
Write a function to sort data✅ Strong:
Write a Python function that takes a list of dictionaries representing users (with 'name', 'age', 'score' fields) and returns them sorted by score in descending order. Include type hints and a docstring.Give the LLM all the information it needs. This includes relevant code, error messages, requirements and background.
❌ Weak:
Fix this bug✅ Strong:
I'm getting a "TypeError: 'NoneType' object is not subscriptable" error on line 45.
Here's the relevant code:
[code snippet]
The function should return a dictionary but sometimes gets None from the API call.
How can I handle this case gracefully?Tell the LLM what perspective to take. This helps shape the tone, depth and approach.
Examples:
- "You are an expert Python developer reviewing code for security vulnerabilities..."
- "Act as a patient teacher explaining React hooks to a beginner..."
- "You are a technical documentation writer creating API reference docs..."
Show the model what you want with 1-3 examples. This can be incredibly powerful for formatting and style.
Example:
Convert these user inputs into structured JSON:
Input: "Meeting with Sarah tomorrow at 3pm"
Output: {"type": "meeting", "person": "Sarah", "time": "15:00", "date": "tomorrow"}
Input: "Remind me to call the dentist on Friday"
Output: {"type": "reminder", "action": "call", "person": "dentist", "date": "Friday"}
Input: "Lunch with the team next Tuesday at noon"
Output:❌ Weak:
Build me a complete web scraper that extracts product data, stores it in a database, and generates visualizations✅ Strong:
Write a function to scrape product names and prices from this HTML structureNow add database insertion using SQLAlchemy for the scraped dataCreate a visualization function that plots price trends over timeBe explicit about how you want your response to be structured.
Examples:
- "Respond only with valid JSON, no explanation"
- "Format your response as a numbered list"
- "Write this as a Git commit message following conventional commits format"
- "Return a Python dictionary with keys: 'status', 'data', 'error'"
Use markers to separate different parts of your prompt, especially when including code, error messages or data.
Example:
Analyze this code for bugs:
```python
[your code here]
```
Focus on: null pointer exceptions, array bounds, and type mismatchesTell the model explicitly what not to do or what limits to expect.
Examples:
- "Keep your response under 100 words"
- "Don't use any external libraries beyond the Python standard library"
- "Avoid using deprecated jQuery methods"
- "Don't include any placeholder or TODO comments—only working code"
Your first prompt rarely works perfectly. Treat it more like a conversation:
Create a REST API endpointAdd input validation for the email fieldUse Pydantic for the validation instead of manual checks- Being Too Polite - LLMs don't have feelings, be direct and clear
- Assuming Context - The model may not accurately remember past conversation. This is especially true in API calls. Always provide full context.
- Asking Multiple Unrelated Questions - Stick to one task per prompt.
- Ignoring Token Limits - Very long prompts or requests for large outputs can hit limits and degrade quality. Prefer smaller chunks.
For more complex reasoning tasks, ask the LLM to show its work and thought process.
Example:
A store has 15 apples. They sell 40% and then receive a shipment that doubles their remaining stock. How many apples do they have now?
Solve this step by step, showing your calculations at each stage.For important decisions (maybe about the direction of your project), ask the model to consider multiple approaches.
Example:
Propose three different architectures for this microservices system, then evaluate the pros and cons of each.LLMs often "reinvent" your code.
Use a prompt like this to ensure it stays consistent the whole time:
For the rest of this conversation, do not change variable names, project architecture, or function signatures unless I explicitly say "ARCHITECTURE RESET".Use a critique loop to refine the solution.
Step 1: Generate the initial solution for this task:
<TASK DESCRIPTION>
Step 2: Critique your own solution, identifying mistakes or improvements.
Step 3: Produce a final improved version, incorporating all critiques.
Return ONLY the final version, along with a short bullet list of what changed.Ensure the output is in a fixed, predefined format.
Return all output in this exact format:
<SPECIFY FORMAT: JSON, markdown table, numbered list, etc.>
Do NOT add explanations or commentary unless asked.- Start with "Create a minimal working example of..."
- Ask for "Quick and dirty" solutions first, optimise later
- Request boilerplate, LLMs often strip it
- Include the full error traceback
- Mention the steps you have already tried
- Ask for multiple potential causes
- "Explain [TECHNOLOGY] as if I am familiar with [TECHNOLOGY YOU KNOW]"
- Tell the LLM about your technical background
- Request minimal examples: "Show me the simplest way to..."
Tired of Alt+Tabing? You can now do all your LLM-powered work from right within your IDE!
This quick guide will focus on VSCode, but setup is quite similar for JetBrains' products (IntelliJ, PyCharm, etc.).
There are many extensions from all your popular LLM providers:
Of the above, the only free options are GitHub Copilot Chat (with expanded quotas if you have redeemed your free Education Pack), BlackBox AI and Gemini Code Assist. You will be able to use your Claude API credits for Claude Code during IC Hack.
Warning
Only install one of these... Performance tanks quickly once you install more than one.
The others all require the respective premium memberships for that provider.
We will therefore be talking only about those free LLM extensions.
To install any of these extensions, simply click the Extensions button on the left hand side to open the extensions panel. Then search for your LLM extension of choice from the list above.
Once you have found your extension of choice, install it by selecting the Install button, then selecting Restart Extensions if prompted.
Once installed, the extensions will appear in your editor.
In the above screenshot, BlackBox AI is the lightning bolt on the left hand side. Gemini is the Gemini logo just below the lightning bolt, and Copilot Chat is the chat logo on the top row.
Tip
At this point, it is recommended that you should restart VS Code to ensure the extensions are properly installed.
With Gemini, you will be asked to sign in when you open the pane. If you activated your GitHub Student Developer Pack, you should also sign in with GitHub Copilot. Both authentication requests will open a window in your browser where you can sign in.
You may notice that you have 3 more icons in the bottom right.
These are GitHub Copilot, BlackBox AI and Gemini, from left to right.
For GitHub Copilot and Gemini, they feature AI code completion. The settings for this can be managed by clicking on their respective icons there.
Important
Even with the Student Developer Pack, GitHub Copilot still has rate limits and quotas. There is a moderate chance you will reach those within the time span of IC Hack. Don't worry, you will not be charged, it will just stop working!
As mentioned above, you can click on the logos of the respective LLMs to open their chat panes.
Since you are in an IDE, these LLMs are configured to already know the background context that you are a software engineer, and that you want some code-related task doing.
To add context, you don't need to copy-paste swathes of code anymore either. For Gemini and Copilot, simply right-click on any file or directory in your current VSCode folder to add it as context.
Use
Add File[/Folder] to Chat for Copilot and Add file[/folder] to Gemini context for Gemini.
For BlackBox, press the + button below the chat input box to add context. You may notice that you can also add Problems. Simply copy-paste your problem (normally a stack trace) after the @problems.
All these AI tools have agent modes. These will complete actions for you (asking consent to actually apply the changes).
Aside from the above exceptions, all other prompt generation recommendations found here apply.
Tip
Always review the changes carefully! LLMs aren't perfect.
MCP stands for Model Context Protocol. An MCP server is a service that exposes tools, data, or capabilities to an AI model in a standardized way, so the model can use external resources safely and predictably.
An MCP server can expose:
- Databases (SQL, NoSQL, vector stores)
- APIs (GitHub, Jira, Stripe, internal services)
- Filesystems or documents
- Custom tools (search, calculations, workflows)
- and etc.
When a user gives the LLM a prompt, the LLM may decide it needs to use external tools. The MCP client (configured alongside the LLM) intercepts this request and communicates with the MCP server, which executes the tool and returns results to the LLM.
The benefit is using one or a couple of standardized MCP servers to access all your tools, instead of writing custom integrations for each tool's unique API and authentication.
Use MCP if:
- Your AI needs live data
- You want multiple tools without prompting them one by one
- You’re building agents, copilots, or dev tools
- You care about permissions and safety
Don't bother using MCP server if:
- It’s a one-off script
- Pure text generation
- No external state
You don’t need to build an MCP server from scratch to get started. There are a lot of MCP servers available that are made to connect different kind of tools.
For a list of available servers and setup instructions, take a look at Popular MCP servers to checkout which server would you like to use and how to install/connect them.