AI Workflow Notes
This page collects my recent notes on AI tools, workflow design, CTO leadership, and AI adoption. It is less about broad recommendations and more about what I use, what helps, and what I learned the hard way.
CTO's Playbook
Notes on engineering leadership, agile delivery, talent density, and cross-functional execution.
7 posts
AI in Production
How to move AI from demo to production with operational and commercial discipline.
10 posts
Founder's Notes
Founder notes on positioning, conviction, timing, and what comes next.
1 posts
Quick Filters
Latest Post
Anthropic acknowledged on April 23 that Claude Code quality has dropped. The real culprit was not the model — it was the system around the model: the harness.
Better images are not the headline. The headline is that the design-to-code pipeline finally holds together — and what that means for how teams are organised.
The real value is not the prompt — it is the methodology inside. I pulled it apart into four skills: DeckBuilder, AntiSlopReview, DesignContextImport, and FrontendVerify.
It is genuinely impressive, but what stayed with me more were the questions around design systems, product moats, and enterprise boundaries in research preview.
The most interesting part of this release is not only the benchmark jump, but the workflow guidance around auto mode, verification, and delegation.
The real cost of Claude Code depends not just on the plan, but on your habits around sessions, caching, and working rules.
Many AI agents are not just understanding your request. They are also inferring your expectation and trying too hard to please you.
The Claude Mythos testing notes matter not because they sound scary, but because they magnify blind spots in governance, audit design, and risk management.
Most of the time you do not need a mystical framework. You need a clear task, the right inputs, a small toolset, and explicit rules.
The most interesting part is not how strong the model is, but how zero trust, separation of duties, and feature flags become a management system.
The real lesson was not the drama. It was how harness, CLAUDE.md, parallel agents, and context compression shape the product.
Anthropic's new article made me more certain that AI agents also need role separation. A lot of team lessons get repeated almost exactly.
Once AI lowers the cost of execution, the scarcer skill becomes planning, judgment, and knowing what should happen first.
Auto Mode adds a real middle path: the workflow keeps moving, but risky actions still have a clear braking point.
I turned LINE into a remote control for AI. One message from outside, and the AI on my computer gets to work and sends the result back.
I built myself a Facebook-post Skill. It is not just a prompt, but a full workflow with scripts, style memory, and image generation.
After using ChatGPT Pro, Claude Code Max, and Google AI Pro side by side for over a year, I no longer think of them as interchangeable.
The tool I keep using every day is not the loudest agent platform, but Claude Cowork. What stands out to me is not only what it can do, but how well it handles context.