<Bold_Prompts> #5

The underrated magic of follow-up prompts

How often do you argue with your LLM?

Welcome to the 5th issue of <Bold Prompts>: the weekly newsletter that sharpens your AI skills, one clever prompt at a time.

This email was supposed to hit your inbox on Tuesday but I was too sick to touch my keyboard. The only thing I could do was watch streams on Twitch.tv (League of Legends MSI tournament anyone?) 

I’m getting better now so thank you very much for being patient with me.

Where were we again?

Ah yes, prompt engineering is the language we use to interact with AI models — and those interactions cover an increasing number of tasks, such as:

• Code generation

• Data analysis

• Image generation 

• Designing AI assistants

Using AI for such tasks will both augment your existing skills and help you unlock new ones. The way to get there is to learn how to write high-quality prompts — and that’s what we do here.

Today we’ll talk about a special aspect of prompting: Follow-up prompts.

Follow-up prompts are short instructions you write to react to your model’s output. You write a first prompt. Your model responds. You follow up. Your model responds to your follow-up. You follow up again and so on.

The goal is to steer your model in a specific direction to:

• Request more details.

• Add new instructions.

• Screen for mistakes

• Change the format of the output

• Explore a different perspective.

• Break down a plan into micro steps.

• Brainstorm wild ideas.

Say you ask your model to write an SQL query and it gets the syntax right but the data wrong. You can follow up with a clarification about which columns to consider and which conditions to apply to select the target data.

The better your follow-up instructions, the more likely your model’s next answer will be correct. Also hey, it’s almost never one follow-up. You often need a few back-and-forth exchanges before you get to a satisfying result.

In fact, many people do more follow-up prompting than prompting. 

They get sucked into the “Just talk to AI” game and start to believe it’s the optimal way of using LLMs. Their argument is you can steer the model however you want with one tiny instruction at a time. But that’s a beginner’s take.

Sure you can get basic tasks done through “a conversation” with your LLM. But if you try to achieve complex tasks like writing complex code from scratch or generating synthetic data, you’ll need a lot of good fortune.

You don’t want to rely on luck to get high-quality outputs from AI models, right? That’s why we spend a lot of time crafting detailed first prompts around here.

See, your first prompt is the most important one. It’s usually the longest because it involves crucial details, clear instructions, and elaborate examples. Picture your first prompt as giving your model a map with a clear path leading to a precise destination.

Detailed prompts will not only improve your output, but they’ll also make your follow-up prompts easier and more effective. Your model is less likely to diverge from the path you want it to follow.

Every clear instruction you add after your first prompt increases your chances of landing on the best possible output.

Ideally, you want to learn how to do both: crafting detailed prompts (and system prompts), and writing sharp follow-ups.

Here we often discuss lengthy detailed prompts.

Today we’ll focus on follow-ups.

We’ll explore a list of follow-up instructions I use on a regular basis. Some may strike you as straightforward. Others will surprise you a bit. But all of them are useful in some way.

We’ll split our follow-up prompts into categories —including ideation, brainstorming, and productivity. 

In practice, the best way to start mastering follow-ups is to memorize three or four of them and try them right away. Every now and then, you can come back to this post and pick up new ones.

Ready to dive in?