During the past six months, I experimented a lot with ChatGPT, developing various tools and products around it.
Now I have lots of interesting insights to share.
So, please welcome — 4 non-obvious ways to use ChatGPT for developers.
1. You can ask ChatGPT to respond in almost any text format — JSON, YAML, XML, Markdown, etc
That gives you lots of opportunities to apply it to interesting use cases.
For example, I want to render a pretty Business Model Canvas.
Basically, it’s an HTML table with some unusual rows and columns.
For that, I can ask ChatGPT to respond strictly in JSON format with defined keys, and then populate a static HTML template with it.
That gives you a lot of freedom. With it, you can generate images, websites, custom components, and so on. Just define the desired scheme and ask GPT to follow it.
2. ChatGPT can draw diagrams (yes, even without plugins)
You can also ask ChatGPT to respond in Mermaid — a diagramming and charting format that uses Markdown-inspired text definitions and a renderer to create and modify complex diagrams.
Note that because ChatGPT is trained on data from 2021, it does not support all the modern Mermaid diagram types. However, it can still be quite useful.
3. Don’t be afraid to use dirty data
Among other things, ChatGPT (and probably all LLMs in general) is very good at seeing through the noise.
Right now I’m building a real-time meeting transcription and summarizing web app.
The time has come to test it during a work consultation. The meeting was in Ukrainian language, but I forgot to change the configuration, and it was set to Russian instead.
The resulting transcription was just terrible, the Russian language Whisper model is very bad at capturing Ukrainian words. But still, I put it into ChatGPT and asked for summarization.
And it did just awesome! It corrected all the mistakes and performed a great wrap-up.
That makes ChatGPT really good for cleaning, summarizing, categorization, and other tasks where there is dirty data involved.
4. Don’t underestimate the power of context. LangChain is useful
At first, I was working with ChatGPT API in simple request-response fashion, without thinking about saving context at all.
But then I met a Product Manager who created a one-page prompt where he described the project he is working on: his current goals, roadmap, tech architecture, etc. Also, he added “You are responding to a colleague who knows all the context” so it would not add unnecessary details.
That has basically made ChatGPT a really useful working assistant for him — now he asks it to ideate, write emails, write documents, requirements, etc.
The thing is, the more information you give to ChatGPT, the more use cases it can help you with, and the more accurate the responses will be. That’s where frameworks like LangChain can really be useful.