Anthropic Claude 3.5 can create icalendar files, so I did this

Original Article ↗ Hacker News Discussion ↗

Practical uses of LLMs for calendars & data extraction

  • Many people use Claude/ChatGPT/Gemini to:
    • Extract dates from PDFs, screenshots, school calendars, banking screenshots, OPML, etc.
    • Generate .ics files, Google Calendar links, CSVs, Markdown checklists.
    • Create recurring events (e.g., kids’ school schedules, conference dates, movie-anniversary calendars).
  • Several describe workflows: first have the model tabulate or summarize dates, manually review, then generate the calendar file or a small script to do it.

Accuracy, hallucinations, and “trust but verify”

  • Users warn that extraction is often “99–99.9%” correct, but small off‑by‑one or missing‑record errors are easy to miss and can be serious in high‑stakes domains.
  • Strong thread around the erosion of trust when LLMs confidently state plausible but wrong “facts”.
  • Suggested mitigations:
    • Use LLMs to write validators or scripts rather than trusting direct outputs.
    • Have multiple models perform the same task and compare.
    • Treat LLM output like junior‑dev work: let it do the bulk, then review.
  • Debate over the phrase “trust but verify”:
    • Some call it an oxymoron; others argue trust has degrees and verification doesn’t negate trust.
    • Alternatives: “assume good faith but check”, “trust does not exclude control”.

Model comparisons and behavior

  • Several report Claude 3.5 as:
    • Better at coding and following instructions than some previous tools.
    • Less prone to blatant hallucinations in some factual and coding tasks.
  • Others note:
    • ChatGPT and GPT‑4o can also generate ICS via text or code interpreter, but may initially refuse or be more awkward.
    • Gemini can do similar extraction with some prompting tricks.
  • Some keep subscriptions to multiple models because each has strengths and different refusal behaviors on sensitive topics.

File formats, interoperability, and ecosystem

  • Strong appreciation for human‑readable, open formats like ICS, CSV, Markdown and OPML because:
    • They’re easy for LLMs to generate and for humans to inspect.
    • They enable ad‑hoc tooling and automation outside proprietary silos.

Community sentiment

  • Many are enthusiastic about LLMs as “PDA‑like” assistants for tedious data entry.
  • Others worry about verification overhead, liability, and the forum veering into product‑promo territory.