Claude

Press m for madewithlove

Moving one overloaded ~/.claude folder into three Claude Code accounts is mostly config: CLAUDE_CONFIG_DIR, the hidden ~/.claude.json, plugin paths and a logout. A small zsh function then asks which account to use, and resume finds the right one on its own.

My blog posts are 100% AI-generated

The final sentences are close to 100% AI-generated. The experiences, opinions, angles and decisions about what to publish are not. A walkthrough of a writing process built on interrogation, rejection and arguing over single lines, and why counting generated words says little about authorship.

The harp and the rain

Ambient music started with a quiet record, a dead speaker and rain. Synthesis has always been about connecting what is already there, and making it used to take weeks. With setup decided once and Claude shaping the pieces, the gap between seeing an idea and running it has closed.

Your model doesn't read like you do

A language model never sees your words. Text is chopped into tokens, each token gets an ID, and each ID gets a spot on a map where meaning is defined by what sits nearby. The fluency you see is built entirely from patterns in numbers.

Visiting hours are over

A line in a notes app promises one day, if I have time, and never says what to do next. A clickable design does: it shows the missing screen, the obvious feature and the person you want to send it to. Ideas should land as prototypes, not as lines in a list.

Watermarking AI

Classifiers stopped spotting AI output once models got good. Watermarks hide a keyed statistical bias in text, images and proteins instead. But marking output is only half the job: without shared keys, nobody else can detect it.

How to build a (basic) agentic pipeline with GitHub

Only three steps between an idea and a preview link need a human: answering questions, approving the plan and checking the result. A GitHub Action routes issue and PR events to a Claude routine that refines, builds test-first and fixes review feedback. The real effort goes into when not to proceed.

I miss a good stick

AI made doing more almost effortless, and that is the problem. When every quiet moment becomes a chance to ask another question or kick off another agent, you trade one deep experience for a dozen forgettable hits. Work at 20x, enjoy life at 1x.

How LLMs actually work (explained like I'm 5)

An LLM does one thing: it predicts the next token. Training nudges that guess towards the right answer billions of times, attention lets it look back at the tokens that matter, and chaining guesses together turns a prediction into an essay.

Giving a codebase a memory

Documentation rots, and what your product does ends up in a few developers' heads. So I stopped writing docs and started extracting them: small, typed, linked notes pulled from the codebase, written for product, with every claim traced back to the line it came from.

How we replaced Fyxer overnight

Fyxer's per-seat price more than doubled, so we built our own meeting notetaker in one night with Claude Code. A blind benchmark on our own bilingual meetings, EU-only storage, and a per-hour cost that puts the whole company below what nine Fyxer seats now cost.

How AI made me twenty times faster at making mistakes

When implementation becomes cheap, what's left? I stopped writing code, ventured into machine learning and 3D game development, and discovered that AI made me remarkably efficient at making mistakes. Fortunately, it also made me much faster at learning from them.

Relive your childhood memories, now with no bugs

As a kid, your skills could only get you so far in a game. Sometimes the game itself was the problem. AI fixed a decades-old goalkeeper bug from a short video, and the real shift is what that means for anyone with an idea.

Worktrunk takes the friction out of Git worktrees

Git worktrees let you run agents in parallel, but every new worktree starts cold, fights over ports and leaves junk behind. Worktrunk wraps worktrees in a few simple commands and a hook system, so one config file gives the whole team warm, isolated environments that clean up after themselves.

Your codebase is part of the prompt

Agents follow the patterns in the surrounding code more reliably than they follow your rules, so every inconsistent naming scheme and half-migrated architecture is a lesson you're teaching them. The fix isn't another rule file. It's deleting the bad example.

Subscribe