Skip to content

Releases

v3.0.0

Choose a tag to compare

@github-actions github-actions released this 17 Nov 10:34
v3.0.0
2bf468b

5 million reviews

v2 read 1.2 million reviews. v3 reads 5,008,684 reviews of 7,786 apps on the App Store and Google Play: 21 categories, the top-free charts of the US, UK, India, Canada, and Australia, Google Play down to rank 200, up to 20 English-language storefronts, the last 24 months.

Of the 2,072,794 reviews rated 3 stars or lower, a fresh random sample of 1,200 was read and labelled one by one:

App Store Google Play
Something just doesn't work (bugs) 35% 35%
Support, refunds, scams 22% 20%
The business behind the app (deliveries, drivers, game balance) 18% 21%
Price, paywall, subscription 20% 18%
Login and account 11% 11%
Ads and popups mid-task 9% 12%
Hard to find or use 9% 7%
A redesign moved or removed something 7% 5%
Slow or laggy 5% 6%
Crashes 5% 4%
AI forced into the app 2% 1%
Broken layouts 1% 1%

About a quarter of complaints are still things a design pass can fix (26% and 24%). Bugs, lag, and crashes together are 41% on both stores. "Purple gradient" came up once in 5 million reviews, from someone asking Copilot to bring theirs back.

4 stars, as promised

A second sample of 300 4-star reviews, 150 per store. 87% of them on the App Store and 61% on Google Play still name a problem, but a different one: a missing feature leads on the App Store (30%), and support and login trouble mostly disappear (2% or less). Thanks to daniel933912 for asking.

New

  • Codex plugin. codex plugin marketplace add erkamyaman/unslopovich, then codex plugin add unslopovich@unslopovich.
  • Update or remove steps for every install path in the README, and a short list of what the skill won't do.
  • Examples. Three before and afters in different styles, and a real run on QuizHub, an open-source quiz app, with the full report and diff.

Research changes

  • The corpus is every app whose collection finished, with no 60/40 store split. Each store was collected to its target, rotating through categories so every category is covered evenly.
  • The population now includes 4-star reviews. The headline numbers stay on 3 stars or lower; 4 stars are reported on their own.

Method, code, and limits: research/README.md. Full tables: research/results/summary.md.

v2.0.0

Choose a tag to compare

@github-actions github-actions released this 15 Nov 12:05
v2.0.0
883e82d

What users actually complain about, with real data this time

v1's numbers were keyword matches on each app's newest App Store reviews, which for a big app is about one week of complaints. v2 redoes the study properly: 1,201,466 reviews of 1,162 apps on the App Store and Google Play, 21 categories, apps from the top, middle, and tail of every chart, up to 20 English-language storefronts, the last 24 months.

Of the 529,434 reviews rated 3 stars or lower, a random sample of 1,200 was read and labelled one by one:

App Store Google Play
Something just doesn't work (bugs) 29% 36%
The business behind the app (deliveries, drivers, game balance) 24% 20%
Support, refunds, scams 22% 18%
Price, paywall, subscription 20% 14%
Login and account 11% 10%
Ads and popups mid-task 10% 13%
A redesign moved or removed something 9% 6%
Hard to find or use 6% 7%
Slow or laggy 5% 7%
Crashes 5% 4%
Broken layouts 2% 2%
AI forced into the app 2% 1%

About a quarter of complaints are things a design pass can fix. Bugs, lag, and crashes together are 35% and 43%. Purple gradients: zero mentions in 1.2 million reviews.

Method, code, and limits: research/README.md. Full tables: research/results/summary.md.

New

  • Six agents. npx unslopovich finds Claude Code, Codex, Cursor, OpenCode, Gemini CLI, and GitHub Copilot and installs for each. --agent, --project, and --dir for the rest. Also a Claude Code plugin marketplace.
  • Web apps and websites. React, Next.js, Vue, Svelte, Astro, Tailwind, plain HTML and CSS. On a phone it checks you didn't build a website; on the web it checks you didn't break one.
  • npx unslopovich detect. 31 deterministic rules over SwiftUI, UIKit, Compose, Android XML, React Native, Flutter, and web source. No model, no network, no dependencies. --json and exit codes for CI.
  • Commands: review, treat, scan, chart, follow-up.
  • Scores out of 24 across six dimensions. Performance is never scored from code: every report says Not profiled until someone profiles.
  • Treatment that doesn't trade one template for another. It calls the case heavy, light, or healthy first, agrees a five-line direction with you before editing, and checks the result against the calm generated default (white cards, gray borders, one blue accent, an icon in a tinted square on every row, a dot beside every label).

Changed

  • The skill lives in skills/unslopovich/, the layout skill installers and Claude Code plugins expect.
  • New before/after image.

Thanks

To everyone who commented on the first post: the "after looks like Claude's default", "colored dots", and "a component pass can't fix perf" comments each changed the skill. And to impeccable, hallmark, and unslop, whose ideas are credited in the README.

1.0.0

Choose a tag to compare

@erkamyaman erkamyaman released this 09 Nov 20:24
792e856

Dr. Unslopovich has read 105,027 App Store reviews.

  • Results from 105,027 reviews of 299 top apps in the US and UK App Stores (keyword matches)
  • Adds Angular Native and NativeScript
  • New checks for labels and numbers that misstate data, and privacy claims the code contradicts
  • Clear severity levels and a new Ask first section
  • Reads the repo's style guide and lets the real brand win over a slop tell
  • Fewer false alarms

npx unslopovich