Thread by @yleflour

🏁 And on with part 6

Let's take a turn and work on the Chrome Extension cause I believe I'll be getting it out way before the VSCode one

I'm not waiting for the web to get promptful so I'm taking maters into my own hands

1/11

As always, #gpt4 is setting up and creating everything for me admirably. Although I see some questionable configs in eslint, prettier and tsconfig if it tries to provide them

Oh 💩 !

2/11

I've been keeping the same chat window to keep as much context as possible. But because of this, I now have to deal with resets

I need to be able to get a new chat window up to speed as quickly as possible ⏱

3/11

Fortunately, I've been updating 2 documents along the way

🔸 .pairprog/system_prompt·md
🔸 .pairprog/project_prompt·md

Those help me get the system up to speed.

"But what's inside" you ask ?

4/11

1️⃣ System prompt acts as the basic AI definer

"You are pAIrprog, a pair programming companion"

"You can use `// COMMAND path/file.ext` in code blocks to provide me with instructions"

"I can use CODE_REVIEW with content"

"Do not apologize" as it's starting to anoy me 🙄

5/11

My system prompt needs to be applied to every response in every project

It is the base configuration of pAIrprog's personality 🤖

If using the API, there is a dedicated field for this

In the webview, you need to remind it from time to time, which gets annoying fast

6/11

2️⃣ Project prompt

My project is now composed of 6 submodules. I don't want each of them to use a different stack

So I defined a project prompt that gets appended to my system prompt: "use pnpm, typescript, ..." and my workspace list of modules with descriptions

Works 👌

7/11

🔢 So now I can boot up a new chat with a clean context in a matter of seconds to work as pAIrprog

With the other commands, I can add the remaining context and give #gpt4 a dir tree or such

But the base context is getting too big...

... too big & too pricey 🤑

8/11

System prompt + project prompt = ~1,500 tokens under #gpt3 calculations

This is 1500 tokens minimum PER API call (more on that later)

~5cts per request in an 8K context
~9cts in a 32K context

And this doesn't take into account #gpt4's completion cost, obviously... 😱

9/11

As of today, I now know that pricing will have an impact on the features of any #GPT4 project

So we may have to think about what a context truly is and how it should be tailored to use cases instead of being generic 🧐

More on that later on

10/11

End of today's 🧵

Thanks for reading. If you enjoyed it, you can subscribe, 🔔, and ❤️ that first tweet. That would help me tremendously

Updates are now dropping everyday weekday at 1PM Paris time / 8AM EDT so be on the lookout

✌️

11/11

https://twitter.com/yleflour/status/1638873319343820801