Shipping an LLM Feature to the Public: 7 Guards That Weren't the Prompt
clementds · reddit · 2026-10-07
A Reddit dev who shipped a notes-to-flashcards LLM feature breaks down the 80% of the work that wasn't prompting: delimiter-tagging user input to stop injected instructions, separate grounding checks per mode, dual rate limits (per-IP plus global daily cap), sanitizing provider errors, forcing output language as a hard bug fix, tolerant JSON parsing, and an editable preview before saving — ranked by when each bit him in production.
More from coding & agent
- Jev: open-source PowerShell module turns AI model output into typed, structured decisions — dfinke · 2026-10-07
- Van Gogh Starry Night town generator adds real-time brushstroke flow fields — simonxxoo · 2026-10-07
- VC's experiment: ChatGPT and Claude flip-flop on retirement investing, showing why vertical AI wins — MartinGTobias · 2026-10-07
- AI reverse-engineers an open-source implementation of the entire Adobe Suite — wen_ragnarok · 2026-10-07
- Teknium fixes Hermes Agent bug that silently dropped lessons for user-owned skills — Teknium · 2026-10-07
- Java Vector API: Writing SIMD Directly Since JDK 16 to Unlock Single-Core Performance — lemire · 2026-10-07