What I Built This Week

Seventeen days, 237 commits across 97 projects — the headline theme this fortnight was using the workspace to solve a real personal problem: finding a job. That effort produced more structured output than anything else this period, but it ran alongside a surprising Pokémon hub expansion and a couple of genuinely new things getting off the ground.

job-hunt — Building a job search engine for one

This started as a folder of markdown files and turned into a small internal SaaS. The trigger was straightforward: applications were going out, rejections were coming back, and there was no clear picture of what was where. The first real intervention was "pick this up, update rejections, I seem to be getting more" — which kicked off a Gmail-synced tracking rebuild that replaced the scattered notes with a single applications.json as the source of truth. From there the system grew: a resume generator that produces tailored one-page PDFs per role, a clone_resume.py script to branch new applications from base variants, and a Wealthsimple-specific daily board watcher that diffs job listings and commits the delta automatically every morning. The watcher has already fired twice with real changes.

The most interesting design call was treating each application as its own folder with a named resume artifact inside, which made the handoff doc (a generated list of apply links mapped to their PDF) genuinely useful rather than another thing to maintain manually. There's also a referral tracking layer — contacts at Shopify, TD, Cohere, and Wealthsimple are logged with status. The hard lesson embedded in a feedback_referral_first.md file: apply first and a friend can't refer you after the fact. That one stung. Currently active daily.

pokemon-web — One hub to rule them all

The Pokémon arcade has been running for a while, but this week it underwent a structural consolidation. What used to be eight separate game repos (Red, Blue, Gold, Silver, Crystal, FireRed, LeafGreen, Emerald) each redirected to a unified hub player with a single shared save-slot system backed by Cloudflare D1. The session that drove most of the June 27 spike started with a request to improve the "View my Pokémon" experience — the party view was fine but there was no good way to browse a full box. What shipped: a gallery with IV insights, sort-by-individual-stat, happiness sorting, group filtering by type, and a cheat menu that lets you drop any obtainable Pokémon (or egg) directly into a save. Passive cloud auto-save was added the following morning — saves fire every 30 seconds so nobody loses progress to a forgotten manual save.

The intervention that shaped the cheat menu was blunt: "ok please stop asking me to write 5-12 lines or small chunks, you are good to make these calls, I like to be the human in the loop playtester." That unlocked a much faster build cycle where the agent made architectural decisions and the human caught bugs in the browser. Real-time cable-link-style battles are architecturally scoped and ready to build — playtest scheduled for tomorrow.

data_explorer — A local sports data warehouse

This project accumulated 23 commits in a single day (June 21) with no sessions logged against it, which means it was running as a background task while other things were happening. The scope is: NBA box scores and play-by-play going back to the full history available from nba_api, PGA Tour hole-by-hole data (7 million scores re-parsed from cached scoreboards with strokes-gained derived from first principles), NFL results from 1966 onward via Spreadspoke, and a scrapekit layer that handles the free-tier scraping so nothing depends on paid APIs. Everything lands in local SQLite databases with a generated SCHEMA.md and a plain-English query layer exposed as a local MCP server — meaning you can ask questions about the data in natural language from inside the agent.

The Polymarket Iran-strike forensics notebook that appeared on June 18 is worth a mention: that's a separate thread entirely, looking at pre-reveal trading patterns. It got filed here because the data infrastructure was convenient. The NHL and baseball pipelines were started in separate scratch sessions this window but haven't been merged in yet — those are the next additions.

dive-channel — Capping at 2k views per Short, now A/B testing hooks

The dive YouTube channel hit a ceiling and the diagnostic session on June 15 was honest about why. The question asked was: "should we be worried that we are capping out at 2k views a short?" The analysis pulled channel stats, compared against other dive channels, and identified the problem — non-megafauna content (schools of fish, coral, swim-throughs) has a hook problem. Sharks and turtles carry themselves. Everything else needs a reason to keep watching in the first three seconds. A batch of July 1–9 Shorts was built as an A/B test specifically on hook devices, with separate upload folders per variant so the results can be cleanly compared against the baseline. The upload and scheduling pipeline is now automated with a daily monitor script and a weekly checkpoint that commits analytics snapshots.

Longform is now in the mix too: a turtle compilation is already scheduled, mantas are next weekend, shark footage is queued for the third longform. The one hard constraint flagged in a provenance file: don't use the Roatan Honduras shark footage — not original material.

The massive spike on June 21 was the data_explorer mass-ingest day — over 6,000 insertions across NBA, NFL, and PGA pipelines in a single automated run. June 27 was the Pokémon hub consolidation, which involved initializing eight separate game repos and redirecting them all in the same session. The pattern otherwise is consistent daily activity from June 14 onward, with almost nothing before that — the window opened mid-search-sprint and stayed busy.

Job-hunt ate a disproportionate share of commits relative to its apparent scope — 34 commits for what is essentially a personal productivity repo — which reflects how much of the agent's time went into structured file management (one folder per application, named PDFs, tracked referral contacts) rather than raw feature work. The Pokémon ecosystem's slice looks small in commit count but the actual churn was enormous in lines touched. Everything else is roughly what you'd expect from a scattershot two-and-a-half weeks.

The MTG Commander table shipped a real multiplayer client with Durable Objects and hidden-hand verification on June 28 — that's the project I didn't have room to cover properly and it deserves its own issue. Next week: playtest the Pokémon cable battles, merge the NHL pipeline, and figure out whether the hook A/B tests on the dive channel actually moved the numbers.

— rickleberry