Live Dev Setup Demo: Terminal, Worktrees, Caddy & Custom Tooling

What We Covered

A live demo session - no slides, just real tools in a real codebase. Streamed on X to ~22 concurrent viewers.

Philosophy & Approach

• Don't marry your tools - stay agnostic, keep configuration and knowledge in markdown files so you can switch tools without losing anything
• AI as cognitive offloading - the real gain isn't the model or benchmarks, it's offloading the mental work of remembering context, searching docs, and holding complex state in your head
• Fault tolerance with AI - don't fight mistakes, invest in guardrails and boundaries instead. Same approach as linters and type systems for human developers
• Session management - managing all work through markdown files for context continuity across AI sessions and tools
• Why Claude Code over Cursor - Cursor's $20/mo model requires aggressive heuristics and abstractions that lead to unpredictable behavior. Claude Code runs in the terminal, doesn't lock you into an IDE

Terminal Ecosystem

• Terminal emulators - Ghostty and Wezterm, chosen for easy configuration. [Ghostty] [Wezterm]
• Shell - zsh with rich prompt showing git branch, status, directory, timestamps. [zsh] [bash] [fish] [oh-my-zsh] [oh-my-posh] [starship]
• The CLI vs GUI mindset - why CLI tools are lighter on system resources, faster to operate, and compound over time

Vim Motions Everywhere

• Vimium - vim navigation in the browser. No mouse needed to click links, scroll, go back/forward. [Vimium]
• Obsidian - markdown editing with full vim motions. [Obsidian]
• Claude Code - has built-in vim mode. [Claude Code]
• Colemak keyboard layout - adopted due to RSI injuries from intensive coding, reduces finger travel. [Colemak]
• Recommendation: start with vim motions in your current editor before switching to neovim

tmux

Terminal multiplexer - run multiple terminal sessions inside a single window, split panes, detach and reattach. [tmux]

• Session management - create named sessions per project, switch between them instantly
• Window and pane splits - horizontal, vertical, navigate by number
• Focus mode - zoom into a single pane when you need to concentrate
• Persistence - sessions survive terminal restarts, can reconnect to remote machines

Neovim

Hyperextensible Vim-based text editor. [Neovim]

• Oil - file management as a text buffer. Create, move, copy, delete files using vim motions instead of a file explorer. "Changed how I work with computers". [Oil.nvim]
• Markdown preview - opens rendered markdown in the browser, live-updating. [markdown-preview.nvim]
• Buffer management - quick save, switch between files

CLI Tools

• atuin - shell history with SQLite database and fuzzy search. Filter by directory, search across sessions. [atuin]
• zoxide - directory jumping. Remembers recent directories, fuzzy matches. "I never remember where things are, and I don't want to". [zoxide]
• bat - better cat with syntax highlighting. [bat]
• fzf - fuzzy finder for everything. [fzf]
• gh - GitHub CLI. Create PRs, view repos, open in browser - all without leaving the terminal. [gh]
• Husky - pre-commit hooks. Catches failures locally before burning CI/CD minutes. Example: saved ~$90K/year for a 5-person team by cutting wasted CI time. [Husky]
• xh - friendlier curl. Cleaner syntax for HTTP requests, less boilerplate than raw curl. [xh]

JSON Processing with jq

• jq - a query language for JSON. "It's a whole world" - endlessly powerful for filtering, transforming, and analyzing JSON from the command line. [jq]
• jqp - interactive jq playground. Pipe JSON into it (cat data.json | jqp) and build queries with live preview of results. [jqp]
• AI tool tip - Claude Code and similar CLI tools support --format json output, which you can pipe straight into jq

Live Demo: Earthquake Data

Fetched ~199 earthquakes from the USGS API (last 24 hours) using xh, saved to a local file, then explored interactively with jqp. Built up from simple to complex - a progression that mirrors real debugging:

1. Orient - what are we looking at? - .metadata - first thing you always do, understand the shape of the data
2. How many events? - .features | length - 199 earthquakes in the last 24 hours
3. Flatten to what matters - .features[].properties | {mag, place, type, sig} - strip away 90% of fields you don't care about
4. Top 10 strongest quakes - sort_by(.mag) | reverse | .[0:10] - 7.1 in Malaysia leading
5. Only significant ones - select(.properties.mag > 4) - severity filter. "Show me only the P0s"
6. Group by alert level - group_by(.alert) | map({alert, count: length}) - 193 null, 6 green. How many events per severity, one line
7. Tsunami warnings - select(.properties.tsunami == 1) - the "oh shit" filter
8. Geographic breakdown - split place strings, group by region, sort by count. California 68, Alaska 53. "This is your 'which microservice is on fire' equivalent"
9. Depth analysis - .features[].geometry.coordinates[2] | {min, max, avg} - min, max, average depth across hundreds of events, instant
10. Full incident report - one query that builds a complete summary object: total events, significant count, tsunami warnings, strongest quake, top 5 hotspot regions. "Full situational report. One command. While someone else is still logging into Datadog"

Why This Matters for Teams

• Save expensive queries locally - fetch logs once from your monitoring service, save to a file, query locally with jq. Avoids burning money on repeated log queries - some services charge hundreds of dollars per query
• Makefile-powered debugging - put common jq queries in a Makefile so junior devs can run complex data analysis without learning jq syntax
• Tool-agnostic debugging - doesn't matter what logging/monitoring service a client uses. If you can curl the endpoint, you can work with the data locally

Bonus Tools

• CleanShot X - screenshot and screen recording tool for macOS. Pin feature keeps floating screenshots on screen - useful for watching logs or reference data while working in the terminal. [CleanShot X]
• tmux copy mode - prefix + [ enters copy mode, turning your terminal pane into a vim buffer. Navigate with vim motions, Space to start selecting, Enter to copy, / to search, n/N for next/previous match. Means you can search, select, and copy from terminal output without touching the mouse

Developer Experience Philosophy

• 7 years consulting with 30+ startups (fintech, cyber, seed to $100M+ raises)
• Focus on creating environments where developers enjoy working - happier devs write fewer bugs and ship faster
• Example: Webpack to Vite migration cutting npm dev from ~60s to ~1s, eliminating phone-scrolling context switches
• Building shared language between design and dev teams to eliminate pixel-pushing friction

Related Reading

• Taming Localhost Ports with Caddy - blog post covering part of the Caddy + worktrees setup