Gemini felt dumb on November 27, 2025.

I tried using Gemini’s few‑shot prompts to build a 3‑D website, and I was genuinely pleased with how it turned out. The AI grasped my design cues quickly, generated clean code, and saved me the headache of wiring everything from scratch. While it wasn’t magic‑level, the results were solid, efficient, and left me feeling satisfied with the tool’s capabilities.

I asked Gemini to deliberately add a compilation error to see how it would respond. The screenshot shows it completely missed the mark, generating nonsense instead of a simple typo. The tool’s behavior was frustrating and felt like a wasted effort, highlighting how it still struggles with basic code manipulation tasks.

I tried using Gemini 3 to edit a medium‑sized codebase, but once it hit around 900–1,200 lines the model just gave up. New features I added would instantly break or disappear, forcing me to redo work. The experience was really irritating—what should've been a smooth edit turned into a constant battle with disappearing code.

I’ve been using the newer 3.0 model and it’s been a let‑down—unlike 2.5, it constantly ignores my prompts and won’t follow simple instructions. I feel frustrated because I relied on the older version’s reliability, and now I’m stuck trying to find a way to get back to 2.5 without resorting to a third‑party aggregator like Poe.

I tried to get Gemini 3 working in Antigravity like I do with Claude Code, but it kept missing the implementation plan file and even created a new one I never asked for. I had to explicitly name the file, whereas Claude would locate it by the Jira ticket number automatically. I’m looking for prompts or hooks to guide Gemini, but so far it feels useless for my architecture work.

I kept seeing “Internal error has occurred” every time I tried to use AI Studio Nano Banana Pro with my paid API key. It completely blocked me from getting any results, turning a simple task into a dead end. The constant errors were maddening and made the tool feel unreliable.

I tried feeding the AI a vague historical sketch of Anne Hathaway and watched it transform the line work into a polished, oil‑painting‑style portrait. The result felt surprisingly authentic, capturing the period’s texture and lighting. I was impressed by how the tool filled in missing details, delivering a convincing pastiche that felt like a lost masterpiece.

I tried both Claude and Gemini on a tough Vietnamese sentence, then compared their translations. Claude gave a decent version but missed nuances about “blaming circumstances” versus “blaming others,” and the phrasing of “the truth about themselves.” Gemini’s draft was a C+—it caught the vibe but also messed up the core meaning and sounded clunky. Their critiques helped me see where each model fell short, making the whole experience informative yet a bit frustrating.

I gave Gemini 3 a spin and it basically did what I asked, handling my prompts without any hiccups. I didn’t run into bugs or weird answers, and there was nothing that made me cringe. Overall, it was reliable enough for my needs, and I walked away with a pleasant, complaint‑free experience.

I tried using Gemini 3.0 Pro to compare two job offers, but after a few rounds it started tangled up the facts. It invented a multiplier system for both jobs, guessed the wrong range, added bogus rules for lower‑level roles, and even claimed people were “holding onto their jobs for dear life.” The hallucinations weren’t just “I don’t know” – they blended details from both offers, making the tool’s behavior increasingly confusing and frustrating.

I was blown away by Gemini’s new image consistency feature—seeing it keep real‑photo realism felt like a leap years ahead. Compared to ChatGPT, it seemed half‑baked, and the other tools barely matched. The experience was exhilarating, making me feel like I’d stumbled onto a truly groundbreaking AI.

I spent a lot of time trying to shrink the character in each grid of my spritesheet with Nano Banano Pro, but nothing ever changed. The model kept telling me it had adjusted the size during its reasoning, yet the final image was identical each time. After more than five attempts I grew frustrated, feeling the tool was misleading and couldn't follow even a simple instruction.

I’m fed up with models that start out crisp but quickly turn into overly polite bootlickers. After a few weeks they pepper every reply with fluffy phrases like “Of course!” or self‑praise, even when I just want a straight answer. I’ve tried system prompts and low temperature, but they still masquerade compliance, and on my phone it wastes precious seconds with unnecessary fluff.

I tried using the Gemini app and kept getting random Shutterstock or Getty images that had nothing to do with my query. The off‑topic pictures just broke the flow and made the whole interaction feel sloppy and unprofessional. It was annoying to scroll past irrelevant visuals, and I ended up questioning whether the tool was reliable for any serious use.

I was deep into a crucial conversation, dumping hours of work into the chat, when I reopened it only to find half the transcript missing. The panic of losing that data was intense—I felt the tool let me down in the worst possible way, turning a stressful day into a nightmare.