Skip to content

Plutonic Rainbows

Chanel

I was a great admirer of Superleggera, a limited-edition Allure that is now sadly discontinued — a short-lived gem that has since taken on almost mythic status among collectors. Bleu de Chanel has probably been my least favourite from their range, lacking the distinctive character I’ve come to expect from the house. I’ve been purchasing from Chanel’s men’s line for over thirty years, during which time I’ve witnessed the evolution of their fragrance portfolio from the classic post-war elegance of Pour Monsieur to the bold, leathered chypre of Antaeus in the 1980s, and the daring, unconventional Égoïste of the early 1990s. These three form what I consider an untouchable trinity — each one a benchmark of its era, yet all timeless in their appeal.

When Allure arrived in the late 1990s, it marked another shift in Chanel’s olfactory direction, offering a smoother, more modern sophistication that still retained depth and character. I was instantly captivated, and I’ve enjoyed many of its variations over the years. By contrast, Bleu de Chanel felt — and still feels — generic, a scent that could have been from almost any brand. That said, I remain hopeful that the new L’Exclusif will embody the heritage, daring, and artistry that have defined Chanel’s greatest men’s fragrances, delivering the masterpiece they truly should have created all along.

Bochum Welt - JIL SANDER

Beautiful new EP that forms a collaboration with fashion house Jil Sander, guided by the creative direction of Simone Bellotti. The album weaves minimalist electronic landscapes with analog warmth, uniting meticulous sound design and a fashion-forward sensibility.

The project is a collaborative EP between Jil Sander and Italian electronic producer Bochum Welt that doubles as Simone Bellotti’s opening statement as the house’s new creative director. Titled JS EP, it was conceived in Hamburg — Jil Sander’s birthplace — and introduced mid-July 2025 with the 'Wanderlust' video, ahead of a boutique/online vinyl release (two cover variants) and wider streaming rollout. Across seven tracks — the record blends ambient/IDM textures and analog warmth with meticulous, minimalist sound design, mirroring the brand’s rigor-meets-emotion aesthetic and signaling Bellotti’s long-term plan to make sound a defining element of the Jil Sander identity.

Blog Update 2

I have completed the first major phase of refactoring my python blog generator, transforming it from a 798-line monolithic script into a modern, modular, and fully backward-compatible system. I introduced clear domain models (Post, Site, Page), a dedicated service layer, repository abstractions, and a Pydantic-validated configuration system, backed by 53 passing unit tests. The build process is now faster, with smart filtering and optimized I/O, while the new CLI offers rich options, structured logging, and robust error handling. This layered architecture, built on solid principles, dramatically improves maintainability, testability, and extensibility, enabling easier feature development, better debugging, and team scalability — without disrupting the existing workflow.

Progress

Making progress with refactoring the blog.

Found a lovely picture of Gail Elliott for Escada, 1989.

Blog Update 1

Preparing to use Claude Code so that I can refactor the blog. Some planning involved.

Planning complete. Work started.

Subagents

I have successfully integrated comprehensive subagent best practices into the Claude Prompt Builder system based on the official Claude Code documentation. The implementation focused on transforming the existing agent recommendation system (v3.6.0) into a sophisticated subagent orchestration framework that follows the key principles of specialized AI assistants with focused responsibilities, restricted tool access, and intelligent delegation patterns. I enhanced the role templates in anthropic_techniques.py to include "PROACTIVELY" language and specific tool restrictions for each agent type, ensuring that security-engineers only have Read/Grep/Edit access for analysis, while implementation agents like python-engineer and frontend-engineer have appropriate modification tools but are restricted to their domain-specific file types.

The core enhancement involved creating a complete subagent orchestration system that automatically detects task complexity and generates detailed delegation plans with primary coordinators, sequential and parallel agent assignments, and clear completion criteria. I updated the pre-execution instructions in enhanced_prompt_builder.py to include comprehensive subagent delegation strategies, modified the prompt assembly process to include orchestration plans in generated outputs, and integrated complexity-based delegation decisions into the enhancement_intelligence.py system. The implementation includes intelligent pattern detection for security and testing requirements, adaptive threshold management based on user feedback, and seamless integration with the existing v3.6.0 architecture. All changes have been tested and verified to work correctly, with the PM2 service successfully restarted to deploy the enhancements to the production environment.

Image Generation Improvements

I recently completed a full modularity audit and architectural overhaul of the FluxBase AI image generation platform. This transformation eliminated 70% of duplicate code across 18 models and reduced blueprint file sizes from hundreds of lines to just a few dozen. By introducing shared base classes, a dedicated service layer, centralized configuration, and automated migration tools, I established a clean, production-ready modular foundation designed to scale and streamline development.

The results were significant: an average 81.4% reduction in blueprint code, faster development cycles, and a consistent pattern applied across the platform. The new architecture not only improved maintainability but also made model creation and feature updates far more efficient. With automated CLI migration tools, comprehensive documentation, and a robust backup system, every model can now be migrated, validated, and deployed with minimal effort and maximum safety.

This modernization effort has positioned FluxBase for long-term growth and stability. The platform now offers a maintainable, scalable foundation that enhances developer productivity, simplifies debugging, and ensures consistent quality. The work involved 44 file changes, over 10,000 lines of new code, and complete backward compatibility, ensuring the transition was both seamless and future-proof.

Prompt Builder Improvements

I successfully enhanced the Claude Prompt Builder v3.5.1 with several critical improvements for optimal Claude Code integration. First, I added a quality control instruction that now appears at the end of every generated prompt (where it logically belongs, not awkwardly in the middle like a party guest who arrives during dessert).

I then completely reorganised the prompt assembly order from a confusing jumble into a logical flow: setup → role → context → task → approach → verification - because even AI needs a proper workflow. Along the way, I discovered the system was hilariously misidentifying debugging tasks as code reviews just because the word "reviewing" appeared in the enhanced text (apparently "reviewing the CSS" meant you wanted a formal code review, not bug fixing), and it was detecting the month "December" hidden inside the word "implement" — which I fixed with proper word boundaries before it started finding "May" in "maybe" and causing temporal chaos.

I also finally transformed those ugly Python object representations such as (ContextParameters(...)) into clean, readable bullet points, because nobody wants to read raw Python objects in their prompts, not even Python developers.

All sections now have clear headers, examples show proper Before/After formatting, and the quality control reminder sits politely at the end where it belongs, ready to ensure everything was done correctly and efficiently.

Beauty in the Beast

My first pressing arrived yesterday, and I was a little concerned that the last track did not rip correctly in XLD. After asking ChatGPT, I learned that you can slow the ripping speed. Lowering the drive speed reduces the disc’s linear velocity at the outer edge, giving the laser more time to track the pits accurately and reducing jitter. This improves the signal-to-noise ratio, enabling the error correction system to recover data that would be unreadable at higher speeds, resulting in cleaner reads and fewer uncorrectable errors in the disc’s most error-prone area.

Chat GPT-5

Released yesterday. I still don't have access. Seems like it has not met with universal acclaim. I think they over-hyped it. If a product is that good, then just let it speak for itself.