Shipping side projects at the intersection of AI, productivity, and my personal interests.
- 🏆 Retrodict: Highest-scoring ARC-AGI-3 agent — 99.86% RHAE, all 25 games solved, 5.5x fewer tokens
- 🔗 ThinHarness: A minimal, opinionated agent harness — focused scope, readable core, easy to fork
- ⚔️ Deckfront (in progress): Building a digital board game that combines deck building and tactical skirmish
- 🧠 OPINIONS.md Agent: Agent that evolves my living beliefs file from my weekly reading highlights
- 📝 Obsidian Drift: Native version control + accumulated IDE-like diff view for Obsidian (750+ downloads)
- 📚 Kindle Storyteller: iOS app for generating on-demand, live-syncing audiobook snippets using AI
- 🎨 revealjs-skill: Coding agent skill for making Reveal.js presentations (300+ stars)
- 🚪 AI Gate: Contributor to an Electron app for using ChatGPT, Claude, and Gemini in parallel, tmux-style workspaces
- 📸 stylegen: CLI for generating images with Gemini/GPT Image 2, supporting style references and parallel batch generation
- 💪 LastLift: iOS app for tracking workout consistency across exercise groups, built with SwiftUI and SwiftData
- 📬 reader-notify: Cloudflare Worker that sends Pushover notifications for new Readwise Reader feed docs
| Type | Title | Date |
|---|---|---|
| Build | how I accidentally got the top score on ARC-AGI-3 with 5.5x fewer tokens | Jul 29, 2026 |
| Analysis | An LLM wiki won't change your life | May 3, 2026 |
| Build | I reverse-engineered Kindle to build on-demand AI audiobooks | Jan 25, 2026 |
| Analysis | AI will never master PowerPoint | Dec 16, 2025 |
| Analysis | Thoughts on the AI bubble | Nov 13, 2025 |
Setup
- Pi with gpt-5.6-sol on high is my daily driver, although I also use Claude Code with Fable 5 when sub capacity permits; full configuration and skills in my dotfiles repo.
- I use cmux to organize many coding agent sessions; 1-2 workspaces per project, 2-4 coding agent tabs per workspace.
- The first 3–10k lines of a greenfield project should usually be built while paired with a coding agent; most of my side projects live in that phase and don't have a consistent stream of user feedback, so my setup is optimized for parallel in-the-loop work.
Agent Workflow
- First step is always to plan; the document is working context for the agent, not something I review line by line. I confirm the intended behavior, ask the agent to surface decisions it is uncertain about, and use HTML artifacts to understand when it's broad/complex.
- Once the plan is ready, the main agent delegates the work to a implementation subagent, runs my review-panel skill + script to get independent feedback from Codex / Claude Code / GLM, and sends synthesized fixes back to the same subagent.
Agent Tools
- Matt Pocock's improve-codebase-architecture skill to refactor and improve my codebases.
- Exa for web search, Context7 MCP for current technical documentation, and grep.app MCP for real code examples from public GitHub repositories.
papercutis a small bash CLI that lets agents log workflow friction to a shared file; I review the list periodically and fix recurring problems (inspired by this tweet from Steve Ruiz).
- I maintain a living set of things I believe in OPINIONS.md, inspired by Kun Chen's post.
- ryanbbrown.com is my personal hub, including a board game log, a tracked book list, and the tools I use.