A few days ago, I decided to test the Claude.ai chatbot, because I heard it was quite capable when it came to code.
Context
I’ve been regularly testing the capabilities of other chatbots (ChatGPT, Mistral, and Gemini) for several months now, for coding scripts in G’MIC, and unsurprisingly, overall it doesn’t work at all.
There are clearly not enough functional examples of code written in the G’MIC language (which is a bit esoteric, it must be said 
) to expect an LLM to have enough information about this language in its training set, and to be able to write code “fluently” in G’MIC.
The result is often a lot of hallucinations, the use of G’MIC commands that don’t exist, obvious syntax errors, etc. Nothing usable, therefore, for someone who wants to “tinker” a little in G’MIC language without knowing too much about it.
And I’m not even talking about having a functional code on the first try. Even after a lot of iterations and making suggestions to the LLMs, we usually get nowhere: the chatbots add new errors to a script while trying to fix other errors, and so on, in an infinite loop.
Trying Claude for implementing a rotating 3D cube
And there, with Claude, I must say I was quite impressed.
I asked him to generate a G’MIC script that displays a rotating 3D cube. No, it didn’t work at all the first time: there were obvious syntax errors in the first script generated. But unlike other chatbots, Claude was able to correct the errors with a few well-thought-out suggestions. And while making the corrections, he didn’t introduce (too many) new errors. After a few iterations, I actually had working source code for a 3D cube.
And I had fun helping him improve his code until he was able to generate a spinning dice, with rotation and zoom controlled by the user, and inertia in the rotation.
Here is the final G’MIC script he generated:
die_face :
128,128,1,3,240
rectangle. 2,2,125,125,1,0,15,15,15
r=14
if $1==1
circle. 64,64,$r,1,200,200,200
elif $1==2
circle. 32,32,$r,1,200,200,200
circle. 96,96,$r,1,200,200,200
elif $1==3
circle. 32,32,$r,1,200,200,200
circle. 64,64,$r,1,200,200,200
circle. 96,96,$r,1,200,200,200
elif $1==4
circle. 32,32,$r,1,200,200,200
circle. 96,32,$r,1,200,200,200
circle. 32,96,$r,1,200,200,200
circle. 96,96,$r,1,200,200,200
elif $1==5
circle. 32,32,$r,1,200,200,200
circle. 96,32,$r,1,200,200,200
circle. 64,64,$r,1,200,200,200
circle. 32,96,$r,1,200,200,200
circle. 96,96,$r,1,200,200,200
elif $1==6
circle. 32,32,$r,1,200,200,200
circle. 96,32,$r,1,200,200,200
circle. 32,64,$r,1,200,200,200
circle. 96,64,$r,1,200,200,200
circle. 32,96,$r,1,200,200,200
circle. 96,96,$r,1,200,200,200
fi
go :
512,512,1,3,0 => background
w[background] 512,512,0,0,"Dé 3D - trackball"
repeat 6 { die_face {1+$>} } image6cube3d[-6--1] center3d. n3d. *3d. 200 => cube
prev_x=-1
prev_y=-1
prev_y2=-1
zoom=0
iax=0 iay=1 iaz=0
iangle=0
friction=0.95
do
mx,my,mb={*,x,y,b}
if $mb==1
if $prev_x>=0
nx:=($mx-256)/256
ny:=($my-256)/256
ox:=($prev_x-256)/256
oy:=($prev_y-256)/256
nz:=sqrt(max(0,1-$nx*$nx-$ny*$ny))
oz:=sqrt(max(0,1-$ox*$ox-$oy*$oy))
nl:=sqrt($nx*$nx+$ny*$ny+$nz*$nz)
ol:=sqrt($ox*$ox+$oy*$oy+$oz*$oz)
nx:=$nx/$nl ny:=$ny/$nl nz:=$nz/$nl
ox:=$ox/$ol oy:=$oy/$ol oz:=$oz/$ol
dot:=min(1,max(-1,$ox*$nx+$oy*$ny+$oz*$nz))
angle:=180/pi*acos($dot)
ax:=$ny*$oz-$nz*$oy
ay:=$nz*$ox-$nx*$oz
az:=$nx*$oy-$ny*$ox
iax=$ax iay=$ay iaz=$az
iangle=$angle
rotate3d[cube] $ax,$ay,$az,$angle
fi
prev_x=$mx
prev_y=$my
else
prev_x=-1
prev_y=-1
if $iangle>0.1
rotate3d[cube] $iax,$iay,$iaz,$iangle
iangle:=$iangle*$friction
fi
fi
if $mb==2
if $prev_y2>=0
zoom:=$zoom+($my-$prev_y2)*1.5
fi
prev_y2=$my
else
prev_y2=-1
fi
+object3d[background] [cube],50%,50%,$zoom,1,5,1,1
w.
rm.
wait 20
while {*}"&&"!{*,ESC}
w[] 0
And honestly, that’s not so bad. Particularly if we compare to what other chatbots were able to generate during my previous tests!
Well, to be honest, it was mostly a test for fun. I don’t think I’ll ever need Claude to code anything in G’MIC language! But anyway, I imagine that for someone who doesn’t know G’MIC as well as I do, it could still be useful for creating functional scripts after a few iterations.
A G’MIC mode for Emacs.
So I thought to myself: if Claude is capable of creating functional things in such an exotic language, why not trying to get him to implement a G’MIC mode for my favorite text editor, Emacs? (what else? 
).
It’s something I’ve wanted to do for a while, but honestly, seriously learning Lisp just to do this one particular thing didn’t seem worth the time investment.
Well, I finally succeeded with Claude.
Please note: it wasn’t a walk in the park. I had to iterate a lot over several days (I’m limited by the number of prompts because I use the free version of Claude). It was sometimes painful: it added errors when I asked for a new feature, etc. But in the end, it worked. And let’s be clear, at the end, it was much faster than if I had had to learn Lisp and the Emacs API to create an editing mode.
Here is the current version of the G’MIC mode for Emacs:
I use it since yesterday, and I admit I find it very nice: it does automatic coloring, but above all, it automatically applies the correct indentation to lines of G’MIC code, and that is priceless for me.
So, that was my experience using the Claude.ai chatbot.
My impression is the same as always with this kind of chatbot: it can be really useful if you understand what you’re doing and what you’re asking it to do. Without that, there’s a good chance you’ll end up hitting a wall!
But I was still impressed, especially by the difference in its code generation capabilities compared to its competitors, in languages that are not necessarily very popular (G’MIC and Lisp).
