Show HN: I'm a dermatologist and I vibe coded a skin cancer learning app

Original Article ↗ Hacker News Discussion ↗

User experience & learning value

  • Many commenters found the quiz eye‑opening and difficult; initial scores around 40–60% were common, with noticeable improvement after dozens of cases.
  • Several said the app made them more likely to book a dermatologist visit and gave them a clearer mental picture of “worrying” lesions.
  • Others found it anxiety‑inducing (“everything is cancer”) and worried it could trigger hypochondria.
  • UI nitpicks: desire for a fixed number of questions per session, better zoom levels, working menu links, and a Safari mobile rendering glitch.

Image balance, difficulty & base rates

  • Users noticed that a large majority of presented lesions are cancerous; some “won” by just always choosing “concerned.”
  • Many argued for a ~50:50 mix of cancer vs benign, or modes focused on “melanoma vs other brown benign things.”
  • Multiple commenters stressed that in real life, cancer is a tiny fraction of all lesions, so training on a cancer‑heavy dataset may bias people toward over‑calling cancer unless base rates are explicitly explained.
  • Ideas surfaced for more nuanced scoring: heavy penalties for false negatives, lighter ones for false positives, and progressive difficulty.

Education vs diagnosis, risk & liability

  • The creator repeatedly framed the app as patient education, not diagnosis: helping laypeople decide “see a doctor now vs watch and wait.”
  • Another skin cancer specialist countered that many cancers, especially early BCCs and melanomas, are not obvious to patients or non‑specialists, warning against overconfidence from a quiz.
  • Several commenters worried users will treat it as a self‑diagnostic tool; comparisons were made to carefully contextualized printed pamphlets.
  • Discussion highlighted that building an actual diagnostic app is technically feasible but blocked by liability, regulation, and the difficulty of managing false positives/negatives at scale.

Medical insights shared

  • Basal cell carcinomas can resemble pimples or scratches but persist and slowly grow; they’re usually slow and non‑spreading.
  • Classic BCC features: “pearly” surface with rolled edges.
  • Self‑screening advice: look for new, non‑resolving or changing lesions; use serial photos; consider full‑body baseline checks.
  • “Ugly duckling” sign (one mole unlike the others) was mentioned, as well as the ABCDE rule and a list of common benign look‑alikes.

AI & vibe coding meta‑discussion

  • The app was “vibe coded” with an LLM in a few hours (single‑file JS, no backend), sparking extensive debate about:
    • Empowering domain experts vs producing low‑quality “shovelware.”
    • Whether quick LLM‑written prototypes are fine as educational tools but dangerous as medical products.
    • The broader future of AI‑assisted coding, security, and the shrinking need for traditional developers in non‑tech domains.