The Nightmare of Onboarding Legacy Projects
Diving into a legacy project with tens of thousands of lines of code and zero comments is every developer’s nightmare. The predecessor is long gone, and the documentation is sparse. Usually, you’d spend at least 1-2 weeks just reading the flow and understanding how modules connect before daring to write your first line of code.
Previously, the common solution was using RAG (Retrieval-Augmented Generation). However, RAG often retrieves fragmented pieces of code, causing the AI to get “confused” due to the lack of a big-picture view. With Gemini 1.5/2.0 Pro, everything has changed. The ability to process 1 to 2 million tokens allows you to feed an entire application’s source code into a single prompt.
I once fed a complex microservices system into Gemini. The result was surprising: the AI took only 45 seconds to accurately explain the data flow that would have taken me a whole day to trace manually.
Why Gemini’s Long Context is a “Heavyweight Weapon”?
Most models like GPT-4 or older versions of Claude have a fairly narrow memory limit (context window). When you copy too many files, they start “forgetting” the beginning or return an overload error. Gemini Pro removes this barrier.
The massive context window offers three major advantages:
- Scan the entire codebase: No need for chunking or complex vector databases. The entire project logic fits right into the AI’s cache.
- Understand end-to-end relationships: The AI recognizes that file A calls a function in file B, and file B affects the database in file C consistently.
- Trace logic bugs: Instead of just checking syntax, it can detect deep business logic errors that are only visible when looking at the system as a whole.
A 3-Step Process to “AI-ify” Your Codebase
To help the AI read code effectively, don’t manually copy-paste every file. We need a more professional and streamlined approach.
Step 1: Package the codebase with Repomix
Repomix is a command-line tool that compresses an entire project into a single text file while preserving the directory structure. It automatically excludes junk folders like node_modules or .git to save tokens.
In your project’s root directory, simply run:
npx repomixAfter a few seconds, you’ll have a repomix-output.txt file. This is the perfect “map” for Large Language Models (LLMs) to read and understand as quickly as possible.
Step 2: Feed data into Google AI Studio
Instead of using the standard Gemini web version, use Google AI Studio. This is a developer playground with full parameter customization and maximum context window support.
- Access Google AI Studio.
- Select the Gemini 1.5 Pro or Gemini 2.0 Flash model for the fastest processing speed.
- Click + (Add Content) and upload the
repomix-output.txtfile.
For a React project with about 50,000 lines of code, it only takes up around 150k – 200k tokens. This is still very small compared to Gemini’s 2-million-token limit.
Step 3: “High-Value” Prompts to Explore the Project
Once the AI has grasped the entire codebase, ask targeted questions. Here are some prompts I often use:
1. Architecture Analysis:
Based on the code, draw a Mermaid diagram describing the data flow from Request to Database. Identify the Design Patterns being used.2. Security Audit:
Acting as a Senior Security Engineer, find vulnerabilities such as SQL Injection or exposed Secret Keys in this codebase. List specific files and lines of code.3. New Feature Guidance:
I want to add a 'Two-Factor Authentication' feature. Which files do I need to modify, and are there any considerations regarding the current logic?Real-world Experience: When AI Finds a Bug in 30 Seconds
I once handled a tough memory leak on a FastAPI system. The app would freeze after running for about 2 hours. After loading 40 code files into Gemini, I asked: “Why is the application’s performance degrading over time?”
In less than a minute, Gemini pointed out a connection pool initialization error inside a loop—something I had missed during manual review. It even rewrote that code snippet using the Singleton pattern, immediately reducing RAM consumption by 40%. It saved me at least half a day of digging through logs.
Important Notes to Avoid Pitfalls
While Gemini is powerful, you still need to follow safety principles:
- Data Security: Never upload code containing real API keys or DB passwords. Ensure Repomix has excluded
.envfiles before compressing. - Verify Results: AI can still hallucinate. Never copy-paste AI code without testing and reviewing it thoroughly.
- Manage Costs: AI Studio has a free tier, but if you’re using the API for large projects, monitor your token usage to avoid shocking bills.
Leveraging Long Context is the fastest way to turn an unfamiliar codebase into your home turf. Instead of spending weeks getting acclimated, you can now start solving problems immediately.
