5 NEW Vibe Coding Repos You Haven't Heard of Yet
A 27-minute tool tour through five GitHub repos that make invisible AI-coding problems visible — architecture, complexity, prompting speed, code quality, and security.
June 17thSean Kochel road-tests Open Design — a 22.4K-star, BYOK, local-first clone of Claude Design — by shipping a landing page, an iOS app, and a desktop chat UI in under fifteen minutes of total prompting.
Open Design achieves professional UI outputs comparable to Claude Design by combining a chosen design system template with structured prompt architecture, making it a free, open-source alternative that doesn't lock you into a single provider.
Open Design is a free, local-first, BYOK clone of Claude Design with 22.4K GitHub stars, plug-in skills, 71 built-in design systems, and media-generation APIs that emit raw HTML. The core method is a two-axis prompt: pick a named design system to fix the aesthetic, then feed a structured spec � landing-page section list, screen-by-screen UX brief, or a screenshot for multimodal translation � drawn from a real PRD rather than vibes. The same engine produced an Anthropic-styled converting landing page, three iOS app screens, and a desktop chat UI inside fifteen minutes of prompting. The takeaway for you: AI-slop output is a prompt-and-system problem, not a model problem, and you can swap aesthetics or platforms by changing inputs instead of tools.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →
States the hook: an open-source clone of Claude Design is free, and after a few days he might not go back. Promises three concrete examples.

GitHub tour: local-first, BYOK, works with any CLI agent (Claude Code, Codex, Cursor, Gemini), 22.4K stars, system-prompt-driven, customizable skills, 71 built-in design systems, HTML output, inspired by four upstream projects.

Walks through the showcase: ready-to-copy example prompts; design-system templates with full design.md spec files; built-in media APIs for image/video/audio plug-ins (e.g. OpenAI gpt-image-2).

Picks Anthropic design system. Names the project. Frames the two real problems: AI-slop aesthetics and unstructured pages that don't convert. Argues design-system solves problem one; prompt structure solves problem two.

Pastes an 11-section landing-page outline (hero, social proof, problem, how it works, key benefits, testimonial, use cases, comparison, case study, FAQ, final CTA) plus an executive summary lifted from his app's PRD as context.

Open Design fires its generative-UI clarification round. He answers a few questions, hits send. Highlights the built-in SaaS-landing-page skill and the fact that you can drop in your own copywriting skills.

Five minutes later: a 'pretty professional looking' Margin landing page in the Anthropic editorial style — comparison table, case-study card with prominent outcome, FAQ section. Calls out the structure is convert-shaped.

Same prompt, brutalist tone instead of editorial. Shows the radically different aesthetic the design-system swap produces with no other prompt change.

Same two problems (aesthetic + UX structure). Reuses Anthropic design system. Introduces his custom skill that brainstorms UX paradigms and emits paste-ready prompts. Goes screen-by-screen for iOS.

Daily Inbox home feed, Gap-Closer Feed library recommender, Ad-hoc Log search. Calls out micro-affordances: skip/swap/confirm, BEST FIT badges, library-ranked-for-the-gap pattern.

Re-prompts with 'coaching chat-first interface' philosophy. Result is unrecognizable from variant one — slash commands for log/photo/goals/trends, conversational logging flow. Same backend prompt, different UX north star.

Screenshots the three iOS screens, drops them into a new chat, asks 'mock up a web app version of this core functionality.' Returns a three-pane desktop chat (nav left / canvas middle / progress right) with the slash commands working visually.

Sums up: free, no provider lock-in, matches most Claude Design functionality with more customization. Points to his skill pack in the description and to a prior OpenSpec video for merging generated screens back into a real codebase.
Pairing a structured prompt with a named design system eliminates AI-slop aesthetics and convert-shaped layouts at the same time — before a single line of real code is written.
“An open source developer just cloned Claw Design. It's a 100% free, and after using it for the past few days, I'm not sure I'll go back to the original.”
“Number one, the page looks like shit and has AI slop written all over it. Number two, the page isn't actually structured to convert people into doing something.”
“It has 71 built-in design systems.”
“It's interesting — with slightly different approaches to the UX that we prompted it with, we get wildly different results.”
“So all in all, you can get really different looking things out the other side of this tool, depending on the design systems that you choose, and most importantly, how you actually choose to prompt it.”
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
An open-source dev just cloned Claude Design, slapped a BYOK badge on it, and Sean Kochel says he might not go back. The whole video is a fifteen-minute live demo answering the obvious follow-up: if the thing is free, local-first, and ships with seventy-one design systems baked in, what does it actually feel like to use? He builds three artifacts in a row — a landing page, an iOS app, a desktop chat UI — and the throughline is less 'tool review' than 'here's how to prompt one of these things so the output doesn't read like AI slop.'
Aesthetic is solved by picking a design system. Conversion structure is solved by prompting an explicit section outline. Two separate levers — most people only pull one.
Paste this list into any design-gen tool and you get convert-shaped output instead of a hero + three feature cards. The 'not 10 benefits' constraint is the load-bearing detail.
Sean frames these as the three concrete reasons to switch. Useful frame for any 'why our open-source clone beats the SaaS' product page.
Hold prompt constant, swap the design-system token (Anthropic editorial → brutalist), get a radically different brand feel for free. Reframes 'pick a brand' as a one-token operation.
Same feature set, same design system — but a one-paragraph 'coaching chat-first interface' philosophy line completely re-architects the UI from cards-and-buttons to slash-commands-in-chat. The UX north star is a prompt variable, not a downstream design choice.
“If you actually want access to all the skills that we use to convert these UX approaches into actual screens that we can prompt into tools like Claw Design or Open Design, you can get more information about that in the description below.”
Soft, late, single-line CTA pointing to his paid skill pack. No urgency, no guarantee, no second ask. Then a secondary next-video CTA pointing at his OpenSpec walkthrough. Low-friction but also low-conversion.
00:00
00:03
00:10
00:20
00:23
00:28
00:38
00:48
00:58
01:08
01:18
01:28
01:34
01:39
01:40
01:50
01:54
02:01
02:08
02:08
02:18
02:28
02:38
02:48
02:51
02:58
03:08
03:09
03:16
03:26
03:36
03:44
03:54
04:04
04:14
04:24
04:34
04:36
04:46
04:56
05:06
05:14
05:24
05:34
05:39
05:44
05:44
05:54
06:04
06:14
06:23
06:33
06:43
06:53
07:00
07:10
07:20
07:24
07:31
07:40
07:50
08:00
08:00
08:10
08:15
08:20
08:24
08:29
08:39
08:49
08:53
08:55
09:05
09:12
09:16
09:26
09:36
09:41
09:49
09:55A 27-minute tool tour through five GitHub repos that make invisible AI-coding problems visible — architecture, complexity, prompting speed, code quality, and security.
June 17thSean Kochel wires Mobbin's screen library into Claude Design through a community MCP server, then uses it to generate three UX directions, an onboarding flow copied structurally from a named competitor, and a single mocked-up UI component — all for a lactation-support app he invents on the spot.
August 5thA solo builder scopes, researches, designs, and epics-out a real fitness-tracking app, then hands the implementation to an unattended overnight Claude Code loop with a verification sub-agent watching every phase.
July 24thA one-hour engineering checklist for builders who can ship prototypes but keep breaking production.
June 26thA creator distills Anthropic's own guidance on running efficient Claude Code sessions into six habits across three buckets.
August 18thSean Kochel installs an open-source Claude Code skill, answers a five-question brand interview, and watches it build — and bill him $20 for — a scrollable 3D website.
July 15th