A working manual for prompting Google's models — built on the idea that directness, well-placed structure, and grounded inputs are what turn a vague request into a useful answer. Use the interactive flow below to assemble a prompt for your exact situation.
Pick the kind of work you're doing. The flow narrows toward a tailored template, a checklist of what to include, and the pitfalls most people hit.
Independent of the task. Gemini specifically rewards direct phrasing, question-last ordering, and grounding in real inputs over generic instructions.
Gemini 3 favors directness over verbosity. "Summarize this in 100 words" beats "I'd really appreciate it if you could maybe help me by summarizing this." No hype, no flattery — those phrases actively hurt newer models.
Google's own prompting whitepaper recommends placing your specific question AFTER the data and context, not before. Gemini's attention is best at the end of the prompt — that's where it commits to action.
Google explicitly recommends few-shot over zero-shot. One example calibrates format; 2–3 nail it. This is more emphasized for Gemini than for other models — don't skip it.
"Only use the data in <context>" outperforms "don't make things up" — every time. The Pink Elephant Problem: negative instructions force the model to think about the forbidden thing first.
Live Google Search grounding, 1M+ token context, native multimodal (images, video, audio, code), Workspace integration. If your task involves any of these, lean on them — don't paste paragraphs of stale info Gemini could just fetch.
Gemini 3 defaults to short, efficient answers. If you want long, detailed, or step-by-step output, you have to ask. The reverse of older models where you were trying to get them to shut up.
Click any segment to see what role it plays. Gemini accepts either XML tags OR markdown headings — pick one and stay consistent within a single prompt.
Techniques worth knowing by name. Reach for these when the default prompt isn't getting you there.
Gemini's killer feature: instruct it to use Google Search for current data, then write from real sources. Cite specific statistics from 2025–2026, not stale training data.
Google's whitepaper explicitly prefers few-shot over zero-shot for Gemini. Show 2–5 input/output pairs before the real input — especially powerful for classification, extraction, and style-matching.
Upload images, PDFs, video, or audio directly. Reference them by content in your prompt: "In the screenshot, the error appears at the top right." Native multimodal is Gemini's strength.
Gemini's 1M token window is real, but reasoning still degrades past ~3000 tokens. Place questions AFTER documents. Wrap each input in tags. Don't dump 50 files — curate.
Gemini 3 responds well to explicit workflow structure: 1) Plan with sub-tasks. 2) Execute and track progress. 3) Validate output against the original task. 4) Format the answer.
For multi-source research, use Gemini's Deep Research feature instead of a regular prompt. It plans, searches, browses dozens of sources, and produces a structured report.