If you work in a software team, you know the pain. You ask Product Managers or Designers to fill out a detailed Jira template for bugs (Steps to reproduce, Expected, Actual). They do it for a week, and then go back to dropping a raw screenshot in the team chat with the caption: "checkout is broken".
As an engineer, you're left guessing the platform, the state, and the steps.
I got tired of fighting human nature, so I decided to build a bridge between the lazy chat screenshots and our strict Jira boards. Here is how I built a bot that does the translation automatically.
The Tech Stack
I wanted to build this fast, so I went with my go-to stack:
Next.js (App Router) for the API webhooks.
Supabase for keeping track of chat IDs and Jira connections.
Google Gemini Vision API for the heavy lifting (reading the screenshots).
Jira REST API for creating the actual tickets.
How the workflow actually looks now:
A PM finds a bug and forwards a screenshot to our Telegram bot with a minimal caption (e.g., "login button does nothing on iOS").
The Telegram webhook hits my Next.js endpoint.
I pass the image and the caption to Gemini Vision with a strict prompt: "Analyze this UI screenshot. Generate a bug report title, steps to reproduce, expected result, and actual result."
Gemini returns a perfectly formatted JSON.
The backend fires a request to the Jira API, creating a new Bug issue, attaching the original image, and applying the AI-generated text to the description.
The bot replies in the Telegram thread with the BUG-123 link.
Total time: ~15 seconds.
The Result
Devs get their perfectly structured tickets, and non-technical folks don't have to change their habits. The "garbage in, garbage out" problem is completely solved.
I realized this is a universal pain point, so I actually packaged this workflow into a small micro-SaaS called BotBridge (link in my profile).
Have you guys tried using Vision models for QA or triage workflows yet? Would love to hear how other teams are handling the "lazy bug report" problem!
Top comments (0)