experiment 001 · hoshi.crx · in build

Hoshi

A job hunt is a hundred small rejections with no scoreboard. Hoshi is a Chrome extension that gives it one — XP, constellations, and a mascot that celebrates with you.

A scoreboard for the job hunt.

Firegboy

Fire mode activated. Let's make some noise! 🔥

$ wc -l spec.md

123functional requirements

76user stories

$ cat process.log

How this got made — what was specified, what was corrected, and what got rejected along the way.

01 / moodboardpinterest

Before any AI tool opened, a Pinterest board set the taste: hand-drawn sticker art, blob characters with thick outlines, dot eyes, rosy cheeks. The board became the contract — every generated screen got judged against it, and 'reads like an emoji pack' meant rejected.

02 / specifyclaude

The PRD came before the code: 123 functional requirements and 76 user stories written with Claude — the XP economy, constellation-level progression, streaks and badges, calm mode, LinkedIn scanning, AI autofill. Writing gamification as requirements forced the real decisions a moodboard never would: what earns XP, what a level means, what happens on a rejection.

03 / explorefigma make + chatgpt

Figma Make and ChatGPT ran the cheap experiments: fast screen variations and copy alternatives, each judged against the Pinterest contract before anything touched the real build. Most outputs died here — that was the point. Killing a direction in a prompt costs minutes; killing it in code costs a weekend.

04 / correctthe prompting discipline

The build ran on explicit correction, not vibes. System emoji were banned outright — every character is CSS or inline SVG. Two full aesthetic directions were rejected before minimal-sticker-kawaii locked: light Nunito, generous whitespace, color living only in the mascot. And the alumni-email scraper got killed on feasibility and ethics — replaced with LinkedIn alumni search plus drafted connection notes under 280 characters.

prompting notesreference images beat adjectives. negative constraints ('no system emoji') outperform positive ones. one correction per prompt. when the output drifts, restate the whole aesthetic contract — don't patch it.

05 / reflect

The spec is the design. Every hour spent making the PRD precise cut the prompting loops from restarts down to corrections — the AI was only ever as good as the constraint I gave it. Next time I'd write the correction log as I go; it turned out to be the most interesting artifact of the whole build.

built in the workshopThe interactive pieces on this page — the process rails, the count-up stats, the mascot — are themselves built through directed AI sessions, the same way Hoshi was: specified in plain language, corrected when they drifted, and rejected until each earned its place. The world-guessing game (MapGame) on my About page came out of the same workshop.see the workshop

$ git log --decisions

The two calls that shaped what Hoshi is — and what it refused to be.

ethics over automationKilled the alumni-email scraper. Shipped LinkedIn alumni search with drafted intros instead.Auto-harvesting personal emails was a privacy problem gamification couldn't excuse — and cold-email blasts get users flagged, not hired. A warm note someone actually sends beats an automated one they'd regret.

rejectedScraping alumni emails from the school directory for one-click outreach

taste over convenienceBanned system emoji. Every character is hand-built in CSS and SVG.A kawaii mascot rendered in platform emoji reads as an emoji pack, not a product — and it looks different on every OS. Owning the mascot in code kept it on-brand everywhere and gave it real personality to animate.

rejectedUsing Unicode / system emoji for the mascot and its reactions

what exists

Working MV3 extension core: popup tracker, XP, streaks and badges, autofill on major job boards, and the CSS/SVG-only Hoshi mascot.

what's specified

In the PRD, not yet in the build: resume tailoring via Google Docs, cover-letter generation, LinkedIn alumni outreach, Google OAuth.

next step

[Placeholder — name the one concrete next step, e.g. shipping the OAuth flow or the resume-tailoring panel.]

$ cd /lab ← back to all experiments

home

about me

work

play

connect

Hoshi

How Can I Help You?

How Can I Help You?

How Can I Help You?