Yann Leflour Profile picture
Mar 17 37 tweets 16 min read
I already love #GPT4

I love it so much that I'm getting it to turn itself into a fully functional pair programming assistant inside of a #vscode extension

4 hours in, it’s already writing 90% of its own code

To do this, I'm teaching it a new language

1/? 👇
#GPT4 is surely amazing, but its top current limitation is interactivity. Out of the box, #chatgpt4 has no way of directly with outside systems, which is a drag

The thing is, you can actually teach it to

But first, let’s analyze the way we interact with the chat website

When asking for code or shell commands, #GPT4 always responds with a code block

There are a couple of ways I interact with those

1. For commands, select, copy, paste. Line by line

2. For files, I click “copy” button

3. For partial files: select, copy, paste

Learning #1: The code blocks are a good container

#GPT4 is still very talkative, so we can’t waste time analyzing the whole prompt. But code blocks are self-contained and usually focus only on code or commands

Should be great to get precise actions

Learning #2: The copy-paste tool is the quickest way to interact

It is just a single click, so almost like a “run” button. But I didn’t want the extra step of pasting.

So my first goal was to turn this “copy” into “run” as fast as possible

Let’s ask it to set up a NodeJS script that fetches the clipboard data every 500ms. If it changes, trigger a callback with its content

Copy, paste, edit

I now get the clipboard’s content, but we use the clipboard all the time, so it’s gibberish. How do we fix this? 🤔

We don’t have to get very far in order to achieve this. As devs, we solved this issue a long time ago.

As we can’t always rely on file extensions, most file types start with a header. This is why you can determine most file types just with the binary content.

So I need my code blocks to start with a header which will help me identify #GPT4 commands

Let’s use “// PAIRPROG <COMMAND> <PATH>”

It should be specific enough to avoid mistakes

So I'll define a set of commands and teach it through a carefully crafted prompt.


Iterate a bit. Functional enough.

9/? screenshot of a chat gpt co...
Now my #GPT4 chat renders actionable code blocks. It can:

- Tell me to execute commands
- Tell me to write in files

All with a code block

I can copy those blocks and read them easily

But still no magic… 😞

Fortunately, I now have all the tools to get #gpt4 to self program it’s own extension

👶🏻 “Write the parse function tests”

🤖 “Fine with me”

👶🏻 “First write a PARSE_COMMAND_REGEX for your Commands. I need to extract type, content and path”

🤖 “Sure thing”

Tests pass ✅

👶🏻 “Create executeShellCommand”

🤖 “Of course”

👶🏻 “Create executeWriteFile”

🤖 “You bet”

Let’s wire it (myself) in shell.ts which launches the clipboard watcher, parses the copied items, and calls the right execution method

pnpm start

Magic happens… 🪄

12/? Super cool screenshot of ya...
From now on it will code all its missing executions one by one, all I have to do is click copy, refactor a bit (thanks @GitHubCopilot and @p42ai) and test

4 hours in, nice 🔥

Cool cool cool, but not enough…

But wait up !!!

I’m not done, but I really want to talk about what I’m working on in a Twitter thread.

But a thread isn’t a good thread if it doesn’t end with a call to action 😮

I need to spin up a site fast to collect emails

I know how to buy my 1000th domain name in 10s but I need that sweet code

👶🏻 “I need a nextjs static site using tailwind to collect emails, it’ll use send in blue for form collection”

🤖 ”Let’s do a few things”

click, refactor, delete

pnpm dev

Good enough…

15/? Screenshot of an overly bas...
👶🏻 “This site needs a logo, get me the image prompt to generate it”

Wow, truly ugly results

👶🏻 “Just give me your version as a SVG file, I need a vector based logo anyway”

🤖 “Sure thing”

👶🏻 “Add it to the website”

Aaaah, perfect… 👌

16/? 4 very ugly logosScreenshot of pAIrprog.io l...
Fill out a few forms, click a few buttons (including “copy”)…

Fill out more forms for DNS and whatnot

git push

🪄 pairprog.io 🪄

Wait. End of thread but no extension?

It's on it's way, so make sure to follow my #gpt4 adventures and subscribe to that mighty mailing list 😘


I already have tons bonker ideas in mind for this and it w̶o̶n̶'̶t̶ will write itself 🚀

🏁 Let's reset the 🧵 counter and launch with part 2 of my #gpt4 adventures

Time to go back to work on the extension

👶🏻 Write a "pairProg: open chat" command to open a side webview

🤖 Sure

▶️ copy, run

👶🏻 Now use the #gpt4 chat url


Yeah figures 🤦‍♂️

1/19 chatgpt telling the user th...
💡Learning #3: It's really nice to know that #gpt4 has a good understanding of software limitations. Sure, you can force it to bypass them but at least it "knows" its 💩

👶🏻 So just asking to be sure, cause I have my doubts, can we inject JS in an iframe ?

🤖 Well... Sorry to be the bearer some bad news

Well I know what my next prompt will be...

3/19 ChatGPT telling the user th...
👶🏻 Okay let's spin up pairprog-chrome-extension from root


I'm definitely not waiting 2 hours...

4/19 ChatGPT telling the user he...
Let's go back to the extension with GPT 3.5 then

✅ Get it up to date on the rules with the initial prompt

✅ Provide project's structure

✅ Go with the prompt

🤖 ...

Well, that's a shame...

5/19 GPT3.5 failing to answer a ...
💡Learning #4: I'm already feeling helpless if I can't keep on using #gpt4. And it's a weird feeling I hadn't had for a while

It's like going back to pure JS after working with @typescript for a long time

But worse

I'm still not waiting 1.8 hours, so, what should I do 🤔

☝️Actually GPT3.5 is not that bad but it needs way more context

I've been using a single #gpt4 chat for a couple of days, of course a brand new GPT3.5 chat needs to play catch up.

And it actually comes with the bonus of increased prompt parsing and response speed ⚡️

My guess is that the prompt and response speed should be taken into account based on the problem to solve

#gpt4: Incredibly good at complex tasks, requires basic context. But slow as hell

#gpt3: Very fast. But needs more context and to be trained with updated docs

Back to developing...

I'm adding commands left and right for faster interactions in the VSCode extension. GPT3 handles it quite well actually. Thankfully the "code" API hasn't changed much since #gpt3's training.

My initial prompt is now at it's 150th iteration

But how do you write a #gpt3 #gpt4 initial prompt you ask 🤨

Well, let me tell you about "standards" in #LeanEngineering...

For that, I have to thank @Michael_Balle, @regis_medina, @marekkalnik, and a lot of other people (can't tag them all)👇

A standard is about being able to self check and validate something you or someone else delivers

It consist of 3 core parts and an extra 1
🔸Do not

🔸The "Why" is about purpose.

It should tell the person about what you're trying to achieve and why it's important

Prompt: "You can use Commands impact our project by writing code blocks in your response"

🔸The "Checks" are tools to verify the validity of your delivery.

Usually you leave out the precise "how", but with #gpt4 and #gp3 you have to be specific

Prompt: "The first line of any code block must ALWAYS be "// PAIRPROG <ACTION>" no matter the coding language."

🔸The "Do not" are mistakes to be avoided

In the case of #chatgpt, it's usually mistakes in it's interpretation of the rules. Instead of writing more "Check" a "Do not" is easier to verify

Prompt: "Do not a write a code block without a Command"

🔸The "example" speaks for itself

It's usually a bad practice as it gives out too much "how" and humans tend to take it as a strict instruction

But in the case of #gpt4 it's often required to improve it's understanding

I won't write the example here, you've seen plenty

A standard is a way to verify "the best level of quality we know of today"

Most important part is "today"

Which is why I update my initial prompt (mostly with "Do not"s) every time #chatgpt makes a mistake in rule application

Which is why I'm at my 158th iteration 😉

And let me tell you, both #gpt3 and #gpt4 understand standards really really well

And that's why this system will be a key part of my extension 💪

BTW if you want a good introduction to #LeanEngineering I advise you to dig into @regis_medina's "Learning to Scale" book

IMO it's the easiest way to get into Lean Engineering philosophy and tools

I use those everyday and it's likely to pop up again in later threads 🧵

Ok, I think it's enough for Part 2

Don't forget to follow me for part 3 and subscribe to pairprog.io to get news on my extension

See you later 😉


• • •

Missing some Tweet in this thread? You can try to force a refresh

Keep Current with Yann Leflour

Yann Leflour Profile picture

Stay in touch and get notified when new unrolls are available from this author!

Read all threads

This Thread may be Removed Anytime!


Twitter may remove this content at anytime! Save it as PDF for later use!

Try unrolling a thread yourself!

how to unroll video
  1. Follow @ThreadReaderApp to mention us!

  2. From a Twitter thread mention us with a keyword "unroll"
@threadreaderapp unroll

Practice here first or read more on our help page!

Did Thread Reader help you today?

Support us! We are indie developers!

This site is made by just two indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member ($3/month or $30/year) and get exclusive features!

Become Premium

Don't want to be a Premium member but still want to support us?

Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal

Or Donate anonymously using crypto!


0xfe58350B80634f60Fa6Dc149a72b4DFbc17D341E copy


3ATGMxNzCUFzxpMCHL5sWSt4DVtS8UqXpi copy

Thank you for your support!

Follow Us on Twitter!