Report: Gemini silently drops the first two-thirds of saved custom instructions
East_Dimension4778 · reddit · 2026-11-15
An unverified user report claims Gemini's web "Instructions for Gemini" (custom instructions) feature has a serious silent truncation problem:
- UI illusion: the settings panel lets you save dozens of detailed custom instructions and returns a clean success state;
- Silent truncation: the backend inference gateway allegedly enforces an undocumented token/length limit — and it's not a clean tail cutoff; the first two-thirds of the instruction block can be bypassed or discarded outright;
- Zero feedback: no error, warning, or notification; the frontend says saved, but the runtime never receives the instructions.
The author argues users relying on strict formatting, grounding constraints, and custom workflows end up blaming themselves or model regression when the system is quietly eating their prompt, and advises not to trust the native settings panel blindly.
More from Models
- Hyped as AGI, Argon Feels Worse Than Gemini 3.8 Flash in Early Use — ZealousidealNeat2246 · 2026-11-15
- Are AI Leaderboards Still Worth Taking Seriously? Rankings Disagree Daily — Snappy-Domino4017 · 2026-11-15
- NVIDIA open-sources SkillEvaluator, a multi-tier framework for evaluating AI agent skills — NVIDIA · 2026-11-15
- Anthropic says start most coding work on Opus 5.5, not Fable 5.2 — effort beats model size — blaizedsouza · 2026-11-15
- Same savings app prompt: Claude vs ChatGPT vs Figma Make UI showdown — Tegadesigns · 2026-11-15
- Epoch ECI Trend Fit: Anthropic's Latest Model Runs ~5 Months Ahead — peterwildeford · 2026-11-15