# The Compound Effect of Small Automations > [!metadata]- Metadata > **Published:** [[2025-09-01|September 01, 2025]] > **Tags:** #🌐 #automation #terminal #claude-code #ai #productivity #workflow ![[the-compound-effect-of-small-automations.png]] It's 9:15 AM. My phone buzzes with a Perplexity notification: "Today's most important news stories." I didn't search for news. I didn't open an app. An AI agent decided what matters and delivered it to me. This is automation in 2025—not just doing tasks, but deciding what needs doing. Seven years ago, I wrote about [[Can the iPad Pro replace the Laptop in everyday life?|whether the iPad could replace a laptop]]. I've been iPad-first ever since, but the past 18 months forced me back to the Mac for [Cursor IDE](https://cursor.com/). Then I discovered something unexpected: by SSH-ing into my Mac Mini from the iPad's terminal, I could run [Claude Code](https://www.anthropic.com/claude-code), access all my automations, and maintain persistent sessions with [tmux](https://github.com/tmux/tmux/wiki). The iPad became viable again—not as a standalone device, but as a portal to my always-on Mac server. ## The Evolution of Personal Automation Looking back at my [[18 Months of Learning to Build Software with LLMs]], I see three phases that converged into something powerful: **Phase 1: Manual Automation (Pre-2024)** Keyboard Maestro macros, Hazel rules, Apple Shortcuts. Solid, dependable, but rigid. My git automation pushed twice daily whether needed or not. Everything required exact conditions to work. **Phase 2: Building With AI (2024)** LLMs let me *describe* what I wanted. I built a git monitor that reads changed files, understands whether I added features or fixed bugs, and writes meaningful commit messages. Created Obsidian capture systems that [[Building an AI-Enhanced Quick Capture System for Obsidian|enhanced notes automatically]]. Smarter, but I was still the architect. **Phase 3: Delegating to AI (Late 2024)** Claude Code + [MCP servers](https://www.anthropic.com/news/model-context-protocol) changed everything. Instead of writing scripts that follow rules, I now have AI agents that understand intent. I can say "find all mentions of our API refactor from last week" and it knows to search across daily notes, project files, and git commits. The shift from "do exactly this" to "achieve this outcome" was revolutionary. ## Real Automations That Compound Every evening at 11:25 PM, my Obsidian vault commits itself with an AI-generated message like "Enhanced documentation structure and added meeting notes from product sync." I didn't write that. An automation read my changes, understood them, and described them better than I would. My Perplexity agents run twice daily, curating news and stock updates. The prompt is simple: "What are today's most important and widely-discussed news stories?" The AI decides what's important. That's delegation, not automation. Between my Mac Mini and MacBook Air, `auto-config-sync` keeps dotfiles, configs, and scripts synchronized using [Unison](https://www.cis.upenn.edu/~bcpierce/unison/). When I create a new Claude Code slash command, tmux config, or shell alias on one machine, it appears on the other within minutes. Built this because I kept forgetting to sync my productivity tweaks. The `/search` command in my terminal can find anything in my 5,000+ notes in under 2 seconds. It runs multiple [ripgrep](https://github.com/BurntSushi/ripgrep) patterns in parallel with [fd](https://github.com/sharkdp/fd), searching titles, content, and YAML metadata simultaneously. No AI needed—just smart pattern matching. Though I'm now exploring RAG systems for semantic search when exact matches aren't enough. ## Terminal as the Glue The terminal didn't replace my existing automations—it connected them. Keyboard Maestro still handles GUI automation. Hazel still watches folders. But now they trigger terminal commands that can think. Tmux made this possible. Without it, I'd have to start fresh Claude Code sessions every time I connected—losing context, reloading MCP servers, remembering what I was doing. Now my "life" session runs 24/7 on my Mac Mini with windows for Obsidian, tasks, and calendar. From my iPad, I tap a [Scriptable](https://scriptable.app/) shortcut that SSH's me directly into my existing session. The context persists—I pick up mid-thought, mid-command, mid-workflow. [Zoxide](https://github.com/ajeetdsouza/zoxide) changed how I navigate. Instead of typing full paths, I just type `z proj` and I'm in my projects folder. It learns where I go and makes educated guesses. Perfect for [[My Experiences and Challenges with ADHD|ADHD]]—I don't have to remember paths, just vague intentions. ## The Compound Effect Each automation enables the next: - Terminal provides the interface - Tmux maintains the context - [Tailscale](https://tailscale.com/) enables remote access - MCP servers connect to services - Claude Code understands intent - Git tracking captures everything My morning routine: Open terminal, attach to "life" session, run `/today` for an overview of tasks, calendar events, emails needing responses, and Linear issues requiring attention. Twelve seconds to full context. The same task with apps would take 3-5 minutes and multiple context switches. ## From Instructions to Intent The fundamental shift: we're moving from giving computers instructions to expressing intent. **Old way**: "Every day at 9 AM, run `git add -A && git commit -m 'Daily backup' && git push`" **New way**: "Keep my notes synchronized with meaningful commit messages" **Old way**: Write complex regex to parse and categorize emails **New way**: "Identify important emails and create tasks for them" This isn't about replacing human thinking. It's about delegating the mechanical parts to systems that never forget, never get distracted, never lose context. For those of us who are [[Neurodivergence|neurodivergent]], this external executive function system is transformative. ## Starting Your Own Journey You don't need everything at once. Start with one friction point: 1. **Identify a repetitive task** that interrupts your flow 2. **Define the outcome** you want (not the steps) 3. **Build incrementally**: Start simple, add intelligence gradually 4. **Let automations compound**: Each one enables the next The goal isn't to automate everything. It's to automate the right things—the ones that drain cognitive energy you need elsewhere. Seven years after asking if the iPad could replace my laptop, I realize I was asking the wrong question. It's not about the device. It's about building a system that understands what you're trying to achieve and helps you get there. The future of productivity isn't faster tools or better apps. It's small automations that compound into something greater—a collaborative intelligence where humans provide intent and machines handle execution. I'm building these little blocks for my life, one at a time. It's time-consuming but doesn't feel like a chore—I'm enjoying the process, and LLMs are helping me craft each piece exactly how I need it.