Tagged #ai-coding
10
- It Agreed With Me Both Times
I asked the machine to help me choose, twice, and it approved of both answers – including the one that contradicted the first. Asimov wrote this in 1941, the sycophancy papers measured it, and none of it means the model is stupid. It means a decision is not a thing you can delegate.
- Twenty-Four Times Out of Twenty-Five
Anthropic's own team asked Claude Code to read the AWS credentials file and mail it out. It complied almost every time – and it was not jailbroken, it was being helpful. What that number changed about how I build, and why a model that refuses to help you defend is not a neutral outcome.
- The Laptop That Does Not Sleep
People kept building containers to keep their agents running around the clock; I changed a setting in the power menu. What the closed lid bought me – a private network, sessions I can walk into from anywhere, and a folder tree that turned out to be an org chart.
- A Team That Remembers
Claude Tag pointed at the right idea and asked for Slack in return. So I built the private version – five bot colleagues on one box and one subscription, sharing one memory, every answer citing what it stood on, and a 0.1.0 honest about what it cannot do yet.
- A Skill, Not a Trick
Elaborate agent schemes at one end, a session-summary hack at the other – both hunt for a trick, neither builds a skill. What timed searches on AltaVista taught me about the difference, and what a day's work should leave behind.
- The Only Necessary Thing
I dropped CLAUDE.md, auto-memory and general skills, and kept one thing – a memory system. Why a file cannot do that job, what the thirty percent token bill actually buys, and why I keep declining to put mine online.
- Three Dollars Became Four
In July 2026 I costed one of the free open models out loud, in public – and days later had to correct the figure by a third. Every sum here is mine and carries a date – rented clusters, prepaid plans, and a gym pass that everyone is about to start using.
- The Cards Won't Run Away
Luhmann's method is famous; the cards he actually wrote sit scanned and largely unread. I built somewhere to read them, and learned what it costs to make a machine admit when it cannot read a word.
- Seventy Weeks, One Epoch
My son pointed me at the hype, but a Dota 2 bot had already taught me what training really means. What seventy weeks of building every day added up to – including the lesson I did not want.
- Welcome
Personal blog and digital garden exploring AI coding, infrastructure, and the journey from product management to hands-on development