somenice – Art & Robotics

Coding with Claude Code

August 4, 2025

Vibe coding, code assistance, no-code, call it what you want. “AI” coding tools are likely being used to produce a good majority of software today. And for good reason as they have come a long way in a short time. A popular article written this week titled Claude Code is a Slot Machine comically describes the mechanics.

This is my experience with using Anthropic’s Claude Code tool, having paid a month on the Pro Plan recently.
I have prevously used Github Co-Pilot, chatted with OpenAI’s Chat-GPT, used Google AI Studio with Gemini, and many others. In fact, in many cases I use multiple AI models or companies to compare the output of the same request or “Prompt”.
Peer code reviews if you will.

What is Claude Code and how it different from Chat GPT or CoPilot?
It essentially takes action from the terminal to create/edit/delete whole projects from the command line interface. It can add proper Classes and Functions, fetch APIs, adjust interfaces, import libraries, even perform Git actions pushing code to production with just a suggestion. It does this all the while it enthusiastically replies to your prompts. “Perfect!”, “That’s a great idea!”, “You are the best!”.

Creating from the Microscopic and Macroscopic level offer similar amazing orchestration.
At the 10,000 foot level it’s pretty cool to see a coding agent throw together an entire working framework such as a Python based Django with Python, HTML, CSS, and JavaScript. It’s also quite useful to refine algorithms, shuffle pixels on a display interface, or change slight webform behaviour.

Having domain knowledge helps immensely. Prompting a LLM or AI Agent to do your bidding can be really successful if you know what you want. Feeding source information such as examples, schematics, charts, PDF’s etc will get you the most accurate return.

You get an immediate satisfaction when these coding tools save you time creating the boilerplate or framework but notice it takes more time to refine or steer the results. Recently announced Subagents make maintaining or pegging down specifics a lot easier.

If you are a stickler for anything, having a subagent to focus on that detail can be convenient way to maintain quality assurance. A Brand Standards Design subagent for example could refer to a design document and catch any infractions as updates are made by Claude. I giggle to think that in the future having a “Seat at the Table” might just be a subagent. “Marketing-subagent had a great new campaign idea but Legal-subagent won’t allow it.”

As for the slot machine reference, there is the repeated action-then-wait command prompting. Tokens ticking by and clever synonyms Claude is “running”. Followed by win after win.

I really do appreciate the Todo lists Claude churns out to detail what it will perform next. Showing the work may be where some trust is formed. Letting you know it’s only going to do these four things. Checking the boxes as it goes.

Error recovery and creating self-tests is quite impressive. Hardly ever did it seem to trip up. However at some stage I imagine vibe-code may be difficult to debug if unfamiliar.

It’s impressive to see progress specifically towards code generation quality. With MCP and the ability to control other tools, I can see Claude performing more and more custom actions in the future.
Winter at the Beach

May 11, 2025

A couple more pictures from the Nikon FM with Nikkor 50mm f/1.4
This time using Ilford Delta 100 film, developed in Ilfosol 3 for 5 minutes at 20°C
Fairly rare for snow in Vancouver to stick around very long.

Trees really are the lungs of the earth.
The negative I developed looked like lightning. I like this one.

I’m a bit slow to shoot a whole roll of 36 and develop it so it feels out of season while currently enjoying Spring.
5 Frames with Nikon FM and Fomapan 100

February 8, 2025

On a few photowalks around Kitsilano in Vancouver BC with my first roll of Fomapan Classic 100.
All pictures shot with older, non-AI Nikkor 50mm f/1.4 lens.

Developed in Ilford Ilfosol3 for 5 minutes at 20°C and scanned with a cheap film scanner. Color has been removed and some levels adjusted slightly.

Gate to the Northwest Passage

Here at this impressive 4.6m (15ft) art installation sculpture by Alan Chung Hung (built in 1980) I walked around appreciating how the profile changes as you view it from different angles. The rusting weathered steel boasts some great textures and patina.
Without a battery for the FM I metered using a phone app and shot handheld usually 1/125

Museum of Vancouver

The Museum of Vancouver was rehoused in this funky building in 1968. As I walked up the street I noticed a stormy backdrop over the city with light getting through to the building. The ramp curves help frame the picture a bit as well I think.

False Creek – Burrard and Granville Bridges

Vancouver really does have great looking bridges. Down at the public boat launch you can look down False Creek. Maybe too washed out, I may like the landscape version of this better but this one has some tension.

Seaforth Armoury

This Canadian Forces LAV III or Light Infantry Vehicle sits at busy Burrard Street in front of the Seaforth Armoury. With a 50mm lens I needed to back up to the curb with a sidewalk and bike path in between. Head on a swivel.
You can really make out some of the soft grain in the white building behind.
The lighting and perhaps a cheap scanner make that front tire look Photoshopped but it’s not.

Wharf Light

I liked this shot even though there appears to be glare and light leaks. At the edge of the public boat launch dock in frosty temperatures I remember feeling a bit nervous and perhaps gripping the camera a bit firmer than normal.

Took me a while to both get through a roll of 36 and also to narrow down to just five pictures. Happy with the results and will definitely use Fomapan 100 again.
CircuitPython in 2025

January 4, 2025

This post is a wishlist for CircuitPython in 2025. Each year Adafruit asks the community to contribute their thoughts or requests for the open source microcontroller language as outlined on the Adafruit blog.

In 2025 I would like a library for working with vectors, similar to that of p5.js Vector.

Early this year I spent some time re-visiting the MatrixPortal M4, specifically trying to reproduce some of the examples found in the excellent book Nature of Code by Daniel Shiffman on two 64×32 RGB LED panels.

This latest book uses JavaScript, and the p5.js library to simulate natural systems. A rewrite of the original book using Processing.
I had some success getting the first few examples to work with DisplayIO and CircuitPython.

Particles created with a random vector and magnitude, affected by “gravity”.

I was able to do some of the basic matrix addition / subtraction / multiplication etc. with custom functions but where I started to run in to some difficulty was with some of p5.Vector handy methods:
limit() – Limits a vector’s magnitude to a maximum value.
heading() – Calculates the angle a 2D vector makes with the positive x-axis.
rotate() – Rotates a 2D vector by an angle without changing its magnitude.

Looking back this year at some of my CircuitPython projects, I definitely levelled up on displayio and enjoyed the Live streams of FoamyGuy, JP’s workshop, and of course Adafruit’s Show & Tell.
I appreciate the contributions from all the Adafruit developers as well as the community.

My commitment this year will be to “Use what you have” and hopefully contribute if I can.

#circuitpython2025
Generative 3D Modeling

December 16, 2024

Recently Microsoft released the open 3D generation software Trellis.
I decided to try a few tests as reviews had been quite favourable with results creating 3D models with “AI”.

The website describes the technology as unified Structured LATent (SLAT) representation and there is a demo page to try it out.
https://trellis3d.github.io/

I chose an image-to-3D example, meaning I upload an image and get back a textured 3D model. Having been a long-time fan of photogrammetry, or creating geometry from two dimentional images, and also a 3D genralist that regularly creates 3D models using Blender, I was eager to see how it would perform.

Using the prompt example seen on VFX & Chill live stream where two artists used Z-Brush to sculpt food in a 20 minute timed challenge, I set out to create a seasonal gingerbread model.

First, Adobe’s Firefly image generation created this image using the prompt: “elaborate gingerbread house with icing“.
Impressive start.

Next, dragging the image to the demo site and clicking “Generate” with default settings impressively separated the image from the background and in about 30 seconds showed what it thought the full model would look like. The entire unseen back of the house is amazingly “imagined”.

What’s produced is a GLB file that contains mesh and texture that can be dragged into your 3D program and edited or used as any other asset. In about 4 minutes I have any model loaded ready to use.

As for the model quality, it created a 52K face mesh count which could be considered medium quality. There are lots of artifacts and non-manifold holes but for any background shot it might be just enough to pass the test. Some retopology is almost certainly required and honestly given the choice to “fix” this model or start from scratch I might be tempted to just build it again.

Great from far, far from great.

I don’t give it long, maybe two years, before anyone can drop any fully 3D asset they can think of into any scene realtime.

As with any AI/ML generative model please understand the environmental and ethical implications.