Gemini 3.1 Pro and Gemini 3.5 Flash are Google's text models with a context window of roughly one million tokens. Both can reason, analyze images, accept PDF documents as input, call tools and functions, search the web, and generate images. Both are available in GPTunneL, where you can switch between them mid-conversation. This guide covers how to get the most out of them with well-built prompts.
Flash or Pro: How to Choose
- Gemini 3.5 Flash is the fast, affordable workhorse. Use it for everyday jobs: summarization, rewriting, data extraction, first drafts, and bulk processing of similar requests.
- Gemini 3.1 Pro is built for deeper analysis. Reach for it when you need to reconcile conflicting sources, dissect a complex document, follow a multi-step chain of reasoning, or produce conclusions you'll be held accountable for.
A practical workflow: run the task on Flash first, and only if the answer feels shallow or the logic wobbles, rerun the exact same prompt on Pro. You save both time and tokens.
What to Keep in Mind While Prompting
- A million tokens of context. Entire books, a project knowledge base, or a full correspondence archive fit into one window. Don't slice documents into fragments — load them whole and refer to sections by name. Put instructions at the top of the prompt and repeat the question at the end, so the model doesn't lose the task in the middle of a long text.
- PDF input. The model reads PDFs directly — reports, contracts, slide decks. Don't retell a document in your own words; attach the file and ask your question.
- Multimodality. Screenshots, diagrams, and photos of documents go into the same prompt as text. Say exactly what to look at in the image, or you'll get a generic description instead of an answer.
- Tools and web search. When the answer depends on fresh data, explicitly ask the model to verify it with search and cite sources.
Techniques with Examples
Request Structure
Keep a prompt in four blocks: role → task → material → output format.
You are a financial analyst. Review the attached report
and identify the expense lines that grew the most.
Format: a table "line item — change — likely cause",
followed by three one-sentence takeaways.
Working Through a Long Document in Stages
Even with a million tokens, complex analysis works best in three steps, each as a separate message: extraction → normalization → conclusions.
Step 1: List every obligation of each party mentioned in the document, verbatim.
Step 2: Organize them into a table: party, obligation, deadline, penalty for breach.
Step 3: Based on the table, name the three riskiest clauses and explain why.
Each step then builds on a verifiable intermediate result instead of the model's "memory" of the text.
Few-Shot: Show Examples
When the exact output format matters, two or three examples beat a long description.
Convert reviews to JSON. Examples:
Review: "Fast delivery, but the box was crushed" → {"tone": "mixed", "topics": ["delivery", "packaging"]}
Review: "Everything was great!" → {"tone": "positive", "topics": []}
Now process: "Waited three weeks, support never replied"
Prompts for PDFs and Images
Point at what to look for and what to return — the model sees the whole file but doesn't know what matters to you.
In the attached PDF, find the limitation-of-liability section.
Quote it verbatim and explain in plain language what it means for the client.
The screenshot shows an analytics chart. Name the metric, the time period,
any anomalous points, and give one hypothesis for each anomaly.
Ready-Made Prompts — Copy and Adapt
- Book digest: "Read the attached text in full. Produce a chapter-by-chapter summary: 3–5 key points per chapter, then a one-paragraph overall conclusion and a list of 10 core ideas."
- Contract review: "You are the client's lawyer. Find the clauses in the attached PDF that disadvantage the client. For each: quote, risk, suggested rewording."
- Knowledge base: "Below is our project's entire internal documentation. Answer my questions using only these documents; if the answer isn't there, say so instead of guessing."
- Fresh data: "Search the web for current information on [topic]. List your sources and flag any places where they disagree."
- One-minute draft (for Flash): "Write three versions of a short announcement based on these bullet points: neutral, bold, and formal. Up to 400 characters each."
Troubleshooting
- The answer is shallow — switch from Flash to 3.1 Pro and add: "reason step by step before giving the final answer."
- The model lost a detail in a long document — repeat the question at the end of the prompt and name the section: "look for the answer in the warranty chapter."
- The format broke — add a few-shot example of the desired output plus "return only this, with no commentary."
- It invents facts — ask it to quote the source verbatim with its location in the document, or to use web search and cite links.
- It ignores the image — reference it directly: "base your answer on the attached screenshot, not on general knowledge."
Detailed documentation for the Gemini family lives at ai.google.dev. The easiest way to compare 3.5 Flash and 3.1 Pro on your own task is GPTunneL: both models in one chat, one balance, no subscription.