← AI Hacker Daily

Edition

05

picks

The slop backlash arrives as tooling, not think-pieces.

The slop backlash arrives as tooling, not think-pieces. Everyone agrees agent output trends generic and agent users trend deskilled; today the fixes are installable things instead of essays — a tutorial engine that makes you do the work, agent skills and design systems that scrub the AI look off your UI, a refactoring layer for agent diffs, an Emacs tutor that refuses to type for you, and, as the mirror, a parody component library of every AI-design trope sitting at the top of HN. We dropped the platform news (Apple's Gemini-powered Siri rebuild, MiMo's 1,000-tok/s launch) and Cognition's FrontierCode — a benchmark that "grades like a tech lead" but that you can't actually run. ## Lathe A Go CLI plus a set of agent skills that turn "teach me X" into a multi-part, source-backed tutorial you work through yourself. You type `/lathe build a 3D slicer in Erlang` inside Claude Code, Cursor, or Codex; the skill generates the curriculum and citations, and the CLI serves it in a local reader at port 4242 with verification steps. Brew tap, curl installer, or `go install`; v0.3.0 shipped two days ago; 388 points on Show HN. The inversion is the point: the model does the curriculum design and source-gathering — where it's strong — and you write the code, because doing it is how you learn. Delete the "explain X like I'm a senior engineer who's new to it" threads you keep re-typing into a chat window. Tradeoff: the author cheerfully calls the tool itself vibecoded, and a generated tutorial inherits whatever its sources got wrong — source-backing makes it auditable, not correct. [link](https://github.com/devenjarvis/lathe)

01

ui-skills

Four drop-in skills for design engineers from ibelick (the prompt-kit and motion-primitives author): `baseline-ui`, `fixing-accessibility`, `fixing-metadata`, `fixing-motion-performance`. Standard skills format, so they load into Claude Code, Cursor, or Codex, and they encode the checklists an agent skips by default — focus states, reduced-motion, OG tags, animation jank. 2.4k stars since January. This is the cheap fix for "the agent shipped a page that looks right and fails accessibility and perf review." Delete the 300-line design-guidelines section of your CLAUDE.md that the model half-reads. Tradeoff: four skills is a checklist, not a design system — it makes output competent rather than distinctive, and you're adopting one designer's taste as your baseline.
github.com/ibelick/ui-skills

02

Uiverse Design

Uiverse — the large community CSS-component library — now sells full design systems built for agent consumption: typography, spacing, color, image treatment, and component rules, each shipping a DESIGN.md file the agent reads before it writes a line of UI. The pitch is literally "de-slop your AI generated websites." Same insight as ui-skills from the other direction: don't prompt taste, hand the agent a spec. Delete the "make it look less like every other AI-generated site" prompt, which has never once worked. Tradeoff: it's paid (a launch discount is running), and you're buying someone else's aesthetic — fine for shipping a credible SaaS page, wrong if design is your differentiator.
www.producthunt.com/products/uiverse-io

03

Command Center

A workflow layer that sits above your coding agents rather than replacing them: a refactoring agent that flags duplicated code, hard-coded values, and security smells in what the agent just wrote; diff walkthroughs ordered by logic instead of alphabetically; feedback agents that iterate without polluting the main context. Works with Claude Code and Codex, and bundles an OpenCode-based agent so it runs out of the box. Free tier, then $9–19/month. This is aimed at the actual bottleneck of agent coding — comprehension and cleanup of large generated diffs, not generation. Delete the ritual of pasting a 2,000-line diff into a fresh chat and asking "any problems with this?" Tradeoff: another subscription and another layer between you and the agent, and the free tier caps walkthroughs and refactor runs — trial it on one real diff before paying.
www.cc.dev/

04

codetutor

An Emacs package that opens a right-side tutor panel, watches your file saves, reviews each diff, and answers questions with full project context — and by design never writes into your files. Backends run locally: `codex exec` in a read-only sandbox, or Pi in print mode. Emacs 28.1+. It's two commits and one star — a prototype, not a product — but it's the most direct answer in today's slate to "I can feel my skills atrophying." The save-review loop is the good idea: feedback on the code you just wrote, instead of code you didn't. Delete the inline-completion plugin in one project and see what comes back. Tradeoff: pre-release, Emacs-only, and read-only-by-design means you do all the typing — which is, of course, the product.
github.com/jaketothepast/codetutor

05

Performative-UI

`npm install performative-ui` — a real, working React 19 component library of AI-era design tropes: MIT-licensed, 30KB, TypeScript, with the stated mission of helping you "signal how oversubscribed your funding round is." It hit 988 points on the HN front page, which tells you how precisely it found the nerve. It's here and not in a links roundup because every trope it parodies is one the tools above exist to scrub out, and seeing them collected and named is genuinely useful as an anti-checklist for your next landing page. Delete nothing — install it in a side project and learn the tropes by name. Tradeoff: the joke has a half-life, and the funniest thing about it is how hard the components are to distinguish from this week's actual launches.
github.com/vorpus/performativeUI

One of these,
every weekday.

Free. Unsubscribe by replying with one word. No tracking pixels in the email.