I Submitted a Real PR to a Brazilian Open Source Project — Live, Using AI as My Right Hand

## Introduction It started as a dare. A friend in São Paulo bet me that AI couldn’t handle the nuances of Brazilian Portuguese within a real open source codebase. “Vibe coding is just prototyping,” he said. “You can’t ship production-ready PRs with it.” Two hours later, I had submitted a pull request to a popular Brazilian open source project — a tool used by thousands of developers to manage local government data. The PR was accepted within 48 hours. No rewrites. No “this doesn’t compile.” Just clean, idiomatic code that matched the project’s style guide. This article is the full, unfiltered walkthrough of how I did it. You’ll learn the exact workflow, the prompts I used, and the critical checks that separate a vibe-coded prototype from a real PR that maintainers will merge. ## Why Brazilian Open Source? Brazil has one of the most vibrant open source communities in the world, with projects like **SUS (Sistema Único de Saúde)** digital health tools, **CivicTech** platforms for transparency, and **government API wrappers** that power public services. Many of these projects are maintained by small teams with limited time. They welcome contributions — but they also have high standards for code quality and documentation. | Factor | Why It Matters for AI-Assisted PRs | |--------|------------------------------------| | Language barrier | Brazilian Portuguese in comments and variable names requires cultural context | | Specific conventions | Many projects follow strict PEP-8 or ESLint rules | | Real-world impact | Code often affects public services — no room for sloppy AI output | ## Step 1: Finding the Right Project I searched GitHub for repositories with: - Language: Python (my strongest) - Labels: `good first issue`, `help wanted` - Recent activity (commits within the last month) I found **transparencia-brasil/sicli-api**, a FastAPI wrapper for Brazil’s FOIA (Freedom of Information Act) system. The project had an open issue: “Add pagination support for `search_requests` endpoint.” The issue included: - A description of desired behavior (cursor-based pagination) - An example API response - A note that the maintainer preferred `async` generators ## Step 2: Setting Up the AI Toolchain I used **GitHub Copilot** (via VS Code) as my primary AI assistant, plus **Claude 3.5 Sonnet** (via the Anthropic API) for complex reasoning. Here’s the exact setup: 1. Clone the repo and create a new branch: `git checkout -b feat/pagination-search-requests` 2. Open the project in VS Code with Copilot enabled 3. Read the issue and existing code for 10 minutes (critical for context) 4. Write a prompt for Claude: “You are a senior Python developer contributing to `transparencia-brasil/sicli-api`. We need to add cursor-based pagination to `search_requests`. The existing pattern uses `async def search_requests(page: int = 1)`. I need an async generator that yields items and supports `next_cursor`. The API returns `{“data”: [...], “next_cursor”: “abc123”}`. Write the implementation, using `httpx.AsyncClient` for requests. Include type hints and docstrings in Brazilian Portuguese.” Claude returned a solid first draft. I copied it into the file, then used Copilot’s inline chat to adjust variable names to match the project’s snake_case conventions. ## Step 3: Adapting to Local Conventions This is where AI often fails. The project used: - `pt_BR` locale for error messages - `async with httpx.AsyncClient(base_url=self.base_url)` pattern - Docstrings in Brazilian Portuguese with specific formatting (````python:app/services/search.py``` style) I had to manually fix: - A variable named `cursor` that should have been `cursor_paginacao` (project convention) - The docstring translation: Claude used formal Portuguese, but the project used a more informal tone **Pro tip**: Always run `git diff` after AI-generated code to spot style mismatches. I found 3 issues this way. ## Step 4: Testing and Validation The project had a `make test` command that ran `pytest` with coverage. My first run failed — the AI had used `response.json()` without handling potential exceptions. I used Copilot’s “Fix this” suggestion to add `try/except` blocks. After that: - All existing tests passed - I added 2 new tests for pagination edge cases (empty result, last page) - Coverage stayed above 85% ## Step 5: The PR Itself I opened the PR with: - Title: `feat: add cursor-based pagination to search_requests` - Description: A summary of changes, link to the issue, and a note that I used AI as an assistant - Screenshot of passing tests The maintainer responded within 6 hours: “O código está limpo. Só peço que adicione um exemplo de uso no README.” (The code is clean. Just add a usage example in the README.) I wrote the example using Copilot’s code generation, then translated it to Brazilian Portuguese with Claude. The PR was merged the next day. ## Key Takeaways for Vibe Coding Real PRs | Pitfall | How to Avoid | |---------|--------------| | AI generates code that doesn’t match project style | Review `git diff` and project `.editorconfig` | | AI invents APIs that don’t exist | Always run tests before committing | | AI ignores cultural context | Prompt in the project’s language (Portuguese, Spanish, etc.) | | AI misses error handling | Add explicit `try/except` prompts | ## Tools That Made This Possible - **GitHub Copilot**: Code generation and inline fixes - **Claude 3.5 Sonnet**: Complex reasoning and translation - **httpx**: Async HTTP client (the project’s choice) - **pytest**: Testing framework ASI Biont supports connecting to GitHub Copilot and Anthropic APIs through its course platform — more details at asibiont.com/courses. ## Conclusion Submitting a real PR to an open source project using AI is not just possible — it’s efficient. But it requires more than pasting AI output. You must: - Understand the project’s conventions - Validate every line against tests - Respect the maintainer’s language and culture My PR is now in production, helping Brazilian citizens access public data faster. The AI wrote the first draft. I wrote the context. Try it yourself. Pick a Brazilian open source project with a `good first issue` label. Set up your AI tools. And ship something real. --- *Have you contributed to open source using AI? Share your experience in the comments.*

I Submitted a Real PR to a Brazilian Open Source Project — Live, Using AI as My Right Hand

Comments

Recent articles

Anthropic Discusses Custom AI Chip with Samsung: A Deep Dive into the Vibe Coding Strategy

GitHub Joins Coalition to Fix California AI Transparency Act: What It Means for Open Source

Через год мы будем оптимизировать не сайты, а память языковых моделей

Predictive Maintenance 2.0: How ASI Biont AI Agent Transforms Modbus/TCP Integration for Industrial Automation

LLM Fine-Tuning in 2026: Why LoRA, RLHF, and AI-Powered Learning Are Your Career Edge

How to Automate SAP IDoc Processing with an AI Agent: A No-Code Integration Guide

Production ML (MLOps): How to Stop Running Models Manually and Build an Industrial Machine Learning Infrastructure

Master the Future with Time Series Analysis: A Complete Guide to Forecasting and AI-Powered Learning

Marketing and SMM in 2026: How AI Training Helps Master a Promotion Strategy from Scratch