10 Prompts That Prove Gemini 3.8 Flash Is Next Level

Link to our newsletter: https://bitbiased.ai/ Gemini 3.8 Flash scored nine points higher than 3.7 on one of Google’s coding benchmarks — but if you prompt it the same way you prompt every other AI model, you may barely notice the upgrade. The biggest changes aren’t about fancy prompt engineering. Gemini 3.8 Flash is specifically better at finishing multi-step jobs, checking its own work, handling massive amounts of context, and operating more like an agent instead of stopping at the first plausible answer. The trick is knowing how to prompt those capabilities directly. Released on September 2nd, 2026, just three weeks after Gemini 3.7 Flash, Google is positioning 3.8 Flash as its intelligent workhorse for coding, long-horizon reasoning, and autonomous agents. The benchmark improvements help explain why: DeepSWE reportedly climbed from roughly 65% to 74%, while TerminalBench jumped from around 82% to 91%. And Google is pricing Gemini 3.8 Flash at under $1 per million input tokens, making the model particularly interesting for workloads where you need AI running repeatedly rather than answering an occasional question. But those improvements don’t show up equally everywhere. In this video, I break down the practical prompts that actually expose what Gemini 3.8 Flash is good at — without complicated prompt frameworks or massive system instructions. We start with coding and one surprisingly important change: asking Gemini to test and fix its work before returning the final result. Instead of simply asking it to write code, you can define “done” as producing something that has already been checked. That plays directly into the model’s improvements on longer, multi-step tasks. Then there’s reasoning. Gemini 3.8 Flash offers low, medium, and high thinking levels, with medium as the default. For problems involving real math, logic, or multi-part decisions, explicitly requesting high effort and step-by-step reasoning can make much more sense than using extra computation on a trivial question. Current information is another important boundary. The script notes a training-data cutoff around March 2026, so for newer developments, asking Gemini to use web search can help ground the response in current sources instead of relying on model memory. That still doesn’t remove the need to verify the sources it provides. We also look at prompting Gemini 3.8 Flash for multi-step planning: mapping dependencies, identifying tasks that require earlier work to be completed, and flagging decisions that need human approval rather than letting the model quietly make every assumption itself. Then there’s context. With a context window exceeding one million tokens, Gemini 3.8 Flash can work with enormous transcripts and documents without forcing you to chop everything into tiny pieces. That opens up useful workflows like extracting action items, owners, and deadlines from an entire long meeting transcript in one pass. Gemini 3.8 Flash is multimodal too, accepting images, video, audio, and PDFs as input. But there’s an important limitation: this model analyzes those inputs and responds with text; it isn’t itself an image, video, or audio generator. And 3.8 isn’t better at everything. On Humanity’s Last Exam, the script notes a score of roughly 45%, essentially unchanged from 3.7. The biggest improvements appear concentrated around coding and structured multi-step work rather than every form of open-ended reasoning. So the real lesson isn’t to make your prompts longer. It’s to define what “finished” means: test the code, check the answer, search for current information when needed, plan the entire task, and use the full context you have available. Try one of your normal Gemini prompts twice — once normally, and once with an explicit instruction to check or test its answer before finishing — and compare what comes back. CHAPTERS 00:00 The Gemini 3.8 Flash Prompting Difference 01:17 What Actually Changed, And Why It Decides Your Prompts 02:49 Getting It To Actually Finish The Job Instead Of Stopping Halfway 04:11 Getting It To Check Its Work Instead Of Guessing 05:17 Getting It To Pull Current Information Instead Of Guessing From Memory 06:19 Getting It To Plan Multi-Step Tasks Instead Of Answering One Piece At A Time 07:23 Getting It To Handle The Big Stuff Without Chopping It Up 08:25 Getting It To Actually Look At What You Hand It 09:25 Where It Still Falls Short 10:13 The Verdict #gemini #gemini38 #googleai #promptengineering #ai