Show HN: I vibecoded a 35k LoC recipe app

App functionality & UX

• Core pitch: hands-free, voice-controlled recipe app without long SEO prose. Several users found the voice interaction “slick” and surprisingly effective (“show me the third one” worked instantly).
• Others liked the simple, ingredients-first layout and considered it one of the better recipe UIs they’ve used.
• Numerous bugs and rough edges reported: Firefox white-screen on generation, slow loading, scroll position not reset when viewing a recipe, broken links with certain non-ASCII titles, intermittent downtime while scaling the Heroku database.
• Minor UX feedback: microphone icon should animate while listening, some searches briefly show results then crash.

AI-generated recipes & images

• All recipes and photos are AI-generated. Many commenters found the images deeply uncanny and unappetizing.
• Users rapidly discovered surreal, obscene, and dangerous recipes: bodily fluids, human/animal body parts, cyanide, cocaine, bombs, radioactive waste, “diarrhea walnuts” at 950K oven temperature, etc.
• Some found this hilarious and treated the app as a toy; others saw it as evidence that AI-generated food content is unsafe and unserious.

Content moderation & safety concerns

• Strong calls to add moderation: block bodily fluids, illegal drugs, explosives, hate speech, and obviously inedible or lethal ingredients.
• Suggestions: run an AI safety filter on completed recipes and only surface “sensible” ones publicly, keep user-specified weirdness private.
• Legal and reputational risk was highlighted (e.g., drug and bomb “recipes”).

Vibecoding practice & 35k LOC debate

• “Vibecoding” is described as letting an LLM build most of the app from high-level prompts, accepting code that “feels right” rather than line-by-line review.
• Many are alarmed by a ~35k LOC codebase for a relatively simple app, calling it “slopcoding” or a maintainability and security red flag (duplication, inconsistent logic, over-verbose React patterns).
• Others argue LOC inflation will become normal: LLMs produce features humans wouldn’t bother to hand-code, and future tools (and LLMs themselves) will maintain this code.

LLM tooling, workflows, and limitations

• Multiple reports that LLMs often silently rewrite unrelated business logic, undo manual changes, or “improve” files far beyond the requested edit; automated tests were later added to catch this.
• Some see full-on vibe coding as chaotic and prefer using LLMs for autocomplete, rubber-ducking, or localized edits.
• Workflow tips shared: narrowly scoped prompts, explicit “don’t touch anything else,” conventions files, tools like Cursor, Windsurf/Claude Code, Aider, and memory banks.

Business, cost, and ethical questions

• Concerns about OpenAI audio API and hosting costs with no monetization; some doubt the app’s viral or revenue potential.
• Broader ethical criticism: flooding the web with AI recipe “slop” degrades the commons and makes finding real, tested recipes harder.
• Counterpoint: many users already rely on LLMs for flexible, on-the-fly cooking guidance; they see this app mainly as a tech demo of how far agentic coding has come.