Gemini (Google)

Prompting Gemini

Current guidance for Gemini 3.x, including 3.8 Flash (new stable default), 3.7 Flash, 3.5 Flash-Lite, 3.1 Pro preview, and multimodal variants.

Which model to choose

Gemini’s current model line is not just “Gemini 3.” Use 3.8 Flash as the latest stable default because Google positions it as the most intelligent Flash for long-horizon coding and agentic work. Use 3.7 Flash as the prior-generation option when compute efficiency matters more than peak scores. Use 3.6 Flash as the older balanced option. 3.5 Flash is now legacy. Use 3.5 Flash-Lite or 3.1 Flash-Lite for high-volume work. Use 3.1 Pro preview only when you are deliberately accepting preview behavior for advanced reasoning or coding.

Specialized Gemini models matter because the GUI/API task may not be text-only. Voice, TTS, image generation (Nano Banana 2 and Pro), and video workflows should route to the matching model family instead of trying to coerce a general chat model through prompt wording. Route vulnerability discovery and patching to 3.8 Flash Cyber only when you hold Fairwind Program access because the standard 3.8 Flash endpoint does not carry the same permissions.

Prompting in Gemini

In Gemini’s web products, model and tool availability can vary by account, region, plan, and workspace policy. Write prompts that name the artifact and workflow you want, especially when asking for work across Google files or multimodal inputs.

Gemini tends to reward concise prompts. If a carried-over prompt has a long persona, repeated warnings, and a scripted reasoning ritual, simplify before adding more instructions. Put bulky context first and the actual instruction at the end. Google’s rollout puts 3.8 Flash in the Gemini app for Pro and Ultra subscribers, so the available model depends on plan.

Prompting through the API

Through the API, use the model endpoint for routing and the thinking controls where available. Do not mix old thinking-budget assumptions with current thinking-level guidance unless the selected model’s docs explicitly call for it.

For long or data-heavy prompts, place the source material first:

<data>{{spreadsheet_export}}</data>

Based on the preceding data, identify the three rows most likely to be duplicates
and explain the matching signal for each.

Current prompting guidance

• Be concise and direct.
• Put the actual question after the context for long prompts.
• Use examples and delimiters when they carry real ambiguity.
• Use the model’s thinking controls for depth instead of asking it to “think harder” in prose.
• Remove stale low-temperature or verbosity assumptions when migrating from older Gemini prompts.
• Match thinking effort to the budget. 3.8 Flash does extra reasoning rounds at higher effort and can spend more tokens to get there, so drop to a lower effort level or stay on 3.7 Flash when cost per call matters more than peak scores.

Examples

GUI prompt:

I uploaded a product requirements doc and three customer-call transcripts. Based on
those files, identify the two requirements that are best supported by customer
evidence and the one requirement that looks like team preference rather than
customer need.

API request skeleton:

{
  "model": "gemini-3.8-flash",
  "contents": [
    {
      "role": "user",
      "parts": [
        {
          "text": "<meeting_notes>{{notes}}</meeting_notes>\n\nBased on the preceding notes, return a concise action list grouped by owner. Use null for the owner if the notes do not identify one."
        }
      ]
    }
  ]
}

Prompt upgrade example:

Stale pattern: “You are a visionary strategist. Think deeply step by step and produce a broad answer.”

Better pattern: “Based on the preceding document, recommend the next decision and list the missing information that would change it.”