Samuel Lawrentz - The frontend Dev - Tech Blog

73 lines of YAML that turned lazygit from "nice git UI" into the place I make every commit.

Cron runs your job and forgets it ever happened. kron is the small Rust tool I wrote that keeps every run's output, exit code and duration one command away.

Taskwarrior holds my tasks, timewarrior holds my hours, and a fork of taskwarrior-tui glues them together with Linear-style pickers, a worked-time column and a pomodoro tab.

No AI chat sidebar, no completion plugin.

My .tmux.conf has 28 commits going back to 2023.

One config file, a ctrl+a prefix, and a few keys that jump between coding agents across every workspace.

I went through every SKILL.md in pstack, HumanLayer's skills and Emil Kowalski's skills, counted what they have in common, and read how they're built.

Everyone is making motion graphics with Opus 5.5 right now.

Openship is an open-source, self-hosted deployment platform, a Vercel you run yourself.

Claude Code 2.1.283 stops at a clean point when your 5-hour usage limit hits mid-task, instead of dying mid-edit.

My one rule for AI-written pull requests, messages and docs: only send it if you'd be happy for the reviewer to read the prompt too.

I measured what my Claude Code plugins, hooks, MCP servers and skills add to every request.

Claude Opus 5.5 landed and Claude Code silently made it the default.

Google open-sourced AX, a declarative orchestrator that runs coding agents as sandboxed tasks with a Task, Workspace, Gateway and Model manifest each.

I parsed a month of my own Claude Code transcripts to see where the prompt cache actually misses. 1448 requests, 17 misses, and 16 of them were me walking away for more than an hour.

Kev is a 0.5B open model that takes typed questions and returns probabilities instead of prose, in one forward pass.

Claude Code 2.1.277 reads AGENTS.md as a fallback when there is no CLAUDE.md.

A PermissionRequest hook that DMs every "Claude wants to run X" to Slack and waits for a thumbs up. 36 lines of Bun, approve from your phone, and the agent never blocks on a terminal you are not looking at.

I spent a week smuggling cookies into a headless Chrome so an agent could read my X bookmarks.

The joke is that telling an agent "do not overengineer" gets you an overengineering prevention framework.

I kept my brainstorm maps in h-m-m for a year.

Someone turned a business book into ten Claude Code skills.

A code reviewer that decides which files to look at and which rules apply before the LLM sees a single line.

Claude Code can now run your plugin or skill against test cases, score the runs, then run the same cases without it and show the delta.

A Rust TUI that opens your existing Obsidian vault with backlinks, a force-directed graph and Vim mode.

A three-question prompt that asks an agent to delete before it simplifies and simplify before it optimizes.

A tmux-for-agents just raised a seed round from Bessemer and YC.

I shipped a WebGL fluid simulation on my homepage and Lighthouse gave me a 41.

Your agent builds a page, and you screenshot it back into the chat.

You can tell where someone is in their career by how they bring you a problem, not how hard it is.

The safe incremental path feels responsible, but when a project has a hard part you keep engineering around, it just delays the failure.

School graded us on solving well-defined problems, then AI got great at exactly that.

SVG dies at a couple thousand points, canvas gets you to about 100k, and your screen only has a thousand pixels anyway.

For years you had to pick between fast containers and safe VMs.

Local CI with gh-signoff, a GitHub CLI extension from Basecamp.

I keep editing code, refreshing, and wondering why nothing changed, only to realize I am staring at the prod tab.

I wore never paying for software like a badge of honor.

The org chart tells you who reports to whom and almost nothing about how work gets done.

I lead a team of engineers by day, then kept forgetting to manage my AI the same way.

Everyone is racing to make AI build for them.

Claude's slide feature makes pretty decks you can't edit or diff.

When facing a massive codebase, skip the manual archaeology.

How to give Claude a persistent memory using Basic Memory MCP, with full Obsidian integration for browsing and editing your AI knowledge graph.

How I automated SSH key management, VPN connection with 1Password MFA, and multi-repo sync on my Linux homelab server using Hermes Agent.

Bun 1.3 ships a built-in SQL client.

Ghostty 1.3 added AppleScript support.

nvim-treesitter got archived on April 3.

Travis wakes up every Sunday, researches trending topics, checks my analytics, generates blog ideas, and texts me a summary.

I've been hand-rolling my Neovim config for years.

Claude Code bills by token.