Make tools, not tasks.
Native Mac and iOS apps, and the agent systems that run the rest of my week. The output is never the durable thing — the written procedure is.
Things I've shipped
Small, native, and fast enough that you stop noticing them.
Render
App Store · v1.1.0Displays HTML and Markdown from inside a Shortcut. Three actions — View HTML, View Markdown, Convert Markdown to HTML — plus PDF export, a share extension, and dark mode that follows the system.
MarkdownEditor
Free · open sourceA Mac-native Markdown editor: vim motions, live preview, KaTeX math, Mermaid diagrams, and export to PDF, single-file HTML, or Word. A 5 MB Swift binary, not a browser booting up. macOS 14+, signed and notarized.
Also built, not shipped: StudentDriver — an iOS permit-hour tracker with sunrise/sunset-based day/night classification, WeatherKit logging, and DMV-style PDF export. Finished enough to use, never submitted to the App Store.
Written procedures, run by agents
Each of these is mostly a document. The model is the engine; the discipline is the asset, and it survives the model getting replaced.
Product Team's Tool Chest
Steering files that take a product team from an idea to deep market research, an Amazon-style PRFAQ, a PRD, and a clickable prototype — in a single working session. Research before PRFAQ, PRFAQ before requirements, a check at every seam. Runs in Kiro, Claude Code, and Cursor. Open sourced through AWS Samples.
ceremony-clip
Hand it a two-hour graduation recording and a name. It finds the name being called, confirms the right kid is actually on screen, cuts a ~35-second clip, adds a title card, and scores it with a music model running on the same Mac — then re-transcribes its own mix to prove the name is still audible through the strings. Nothing uploaded.
image-kit
A local image toolkit for agents on the Mac: inspect (dimensions, EXIF/GPS, sharpness, faces), upscale with Real-ESRGAN, repair faces with GFPGAN, cut backgrounds with rembg, and strip metadata before you share. No uploads, no subscriptions.
eval-harness
The scripts from an eval campaign that caught a "perfect" 100% score hiding a 70.8% reality. Repeat-trial runs at production config, per-item flip rates, exact Clopper-Pearson intervals, and episode-level recall — because a perfect score on 30 items is "at least 88.4%, with 95% confidence."
I write about all of this at nateober.com - essays on second-order work and choosing models honestly. Most recent: seventeen models wrote the identical briefing, and thinking mode made them worse.
Things that run in a browser
Not products. They work, and that's the whole claim.
Neon Starship →
A twin-stick arcade shooter in a canvas — waves of neon enemies, power-ups, and a high-score board that has been quietly accumulating other people's initials. Works on a desktop keyboard or a phone thumbstick.
This page builds itself
The site above argues that the written procedure is the durable artifact. It would be a poor argument if this page were maintained by hand — so it isn't.
Compiled from site.toml by build.py at 2026-08-24 04:30 UTC, nightly on cron.
Every word above lives in a 10-entry data file. The numbers that rot — GitHub stars, the App Store version, the high score — are never typed here; they are fetched at build time from the GitHub API, the iTunes lookup API, and the leaderboard's SQLite file.
The previous version of this page said © 2024 for two years and described an app as "coming soon" for one. That is what hand-maintenance buys you. Read the source →