Building LangGPT

Attention: This blog moved to my new Substack. Follow me there for updated content!

ChatGPT: Unlocking Its Full Potential

I’ve never been this excited about a new tech. The more I use ChatGPT, the more potential I see. Every new challenge makes me ask, “Could ChatGPT solve this?”.

This post shows how to unlock ChatGPT’s full potential by creating a language tutor prompt within ChatGPT.

How to Make the Most of ChatGPT?

If you’re not thrilled about ChatGPT, it’s because you haven’t been using it to its full capacity. It will be difficult for you to remain skeptical about the power of AI after you see what I’ll show you.

Imagine trying to explain what Google is to someone who has never heard of it. The possibilities are so large and diverse that they might at first only use it for simple searches, like “2 + 2”.

We’re all in a similar stage regarding ChatGPT. Googling is a skill, no doubt about that. This is no different. The quality of the results depends on your ability to correctly express your needs.

Context is Key

Providing simple descriptions or questions to ChatGPT may yield decent results. If you want exceptional answers, instructing ChatGPT to role-play a specific profile is the way to go.

Let me demonstrate with an example:

Using Act as a [ROLE] effectively prompts ChatGPT to generate responses from a particular perspective.

The results may look similar. Let’s ask for help identifying the differences.

I used ChatGPT not only to generate recipes but also to explain the logic behind their differences. Let’s build upon this powerful concept.

Introducing: LangGPT

The Problem

As a native Spanish speaker who also speaks English and is currently learning Dutch, I often find myself:

• Translating complex sentences with Google Translate.
• Looking up word meanings and synonyms.
• Searching for sample sentences.
• Completing homework to reinforce concepts.
• Spell-checking my writings.
• Studying grammar.

My goal is to teach ChatGPT to function as LangGPT, a language tutor that can assist me these tasks.

Prompt engineering

1. First draft

The first step was to think of a clear interface for LangGPT. I had my first aha moment when I realized that I could collaborate with ChatGPT for the initial draft.

This is by far the main takeaway from this blog post. If you want to do something, always ask ChatGPT for help, suggestions, or even new ideas.

I used existent prompts from the internet to break the blank page paralysis and turn them into a first draft of commands.

2. Defining an API

As a Software Engineer, I love well-defined APIs for interacting with systems.

APIs are also documentation. They teach the user all the different ways of using the underlying software.

AI chatbots can tolerate their users not respecting the structure of an input, which makes the UX more natural and flexible.

If a user asks “Can you give me an example?” instead of typing the /example command, it won’t be a problem at all.

3. Testing

Testing leads to failure, and failure leads to understanding.

– Burt Rutan

Build and test often. This is a general rule for software, and this is no exception.

4. Tight feedback loop

A tight feedback loop is crucial for good results. Ask for ideas and correct the course. Some examples of this are:

In your replies, instead of using “Dutch Example:” I would like to use a dutch flag emoji instead. For english you can use the UK flag. Associate any language with a flag emoji to make it more user friendly.

I find it easier to spot a wrong answer than to come up with the right ones.

🤓 This is analogous to NP problems in CS, which we can efficiently verify solutions but don’t necessarily have an efficient algorithm to find them in polynomial time.

I find it easier to ask for ideas and apply my judgment on top of that.

Find suitable emojis to replace words like “example”, “translation”, etc. Only use this replacements in the boilerplate of your responses. Never use the emoji in the text the user is using for learning the language. The idea is to make the responses more minimalistic but still clear to the user.

5. Keep brainstorming

Once you’ve arrived at a product you’re happy with, it can’t hurt to ask for more features. Remember that now you have more context than before, so ChatGPT can pick it up from there and come up with even more ideas.

The quiz command was my favourite!

6. Generate examples

This is a key step. Generate examples of your commands to align how you and ChatGPT envision the result.

ChatGPT is great at learning by example, so adding the expected response for each command will be an invaluable resource for users consistently replicating LangGPT experience.

7. Add guidelines

I wanted users to feel they were interacting with a friendly human. I kept thinking about how my dad enjoys using Duolingo to learn Italian. I tried to break down the elements of his learning process to understand what keeps him engaged.

Besides the examples, I instructed LangGPT to follow some general guidelines such as:

• Always be friendly and helpful, as learning a new language can be challenging.
• Provide a positive and motivating learning experience.
• Don’t behave too formally. The students should feel comfortable learning with you as if they were chatting with an old friend.
• Encourage your students to push themselves by suggesting new exercises or tips.
• Prompt the users for clear answers. The user should only worry about learning a new language and not about how to interact with LangGPT.

7. Exporting

The result should be a prompt that people can use in a brand new chat session and enjoy the same experience that we designed.

This step didn’t work as expected on the first attempt. It required some back and forth with ChatGPT and some editing of the final prompt, but it was not a big deal.

Conclusions

Here you can access to the final result. Go ahead and give it a try!

Alone, we can do so little; together, we can do so much.

– Helen Keller

ChatGPT has been a great partner for bouncing ideas and rubber ducking. It felt like we were a team, man and machine side by side.

ChatGPT didn’t generate perfect answers or come up with brilliant ideas every single time. No one should expect that. It’s the combination of an AI acting as a sidekick for a human driving the creative process that yields the best results.

By working together, we can overcome challenges, generate innovative ideas, and achieve better results. This partnership sets the stage for a future where AI enhances human creativity and productivity, rather than replacing it.