Keep this page open while you work in Gemini 3.1. It has the prompt formula that makes answers sharper, plus the exact modifiers and trigger phrases that switch on Google Search grounding, Workspace files, Deep Research, and image generation. Every table is copy-paste ready. For finished prompts built on these patterns, see the best Gemini prompts and Gemini prompt templates.
The ICC formula
Almost every strong Gemini prompt has three parts in the same order: Intent, Context, Constraints. State what you want, give Gemini what it needs to know, then set the rules for the output.
| Part | What to write |
|---|---|
| Intent | The task in one plain sentence — what you want produced and why. |
| Context | Background, audience, and any files or data Gemini should use. |
| Constraints | Format, length, tone, must-haves, and what to avoid. |
Here is the same formula with each part labeled, so you can see how they slot together:
[Intent] Write a launch announcement email for our new analytics dashboard.
[Context] Audience is existing customers on the Pro plan. The dashboard adds real-time charts and CSV export. Tone should match our friendly, plain-spoken brand voice.
[Constraints] Under 150 words. One clear call to action ("View your dashboard"). No hype words, no exclamation marks. Return only the email body with a subject line above it.You do not need the literal bracket labels once you are used to it — just keep the three ideas in that order. Grab a role from the next table, a format from the one after, and drop them into the Constraints line. For deeper walkthroughs, see the best Gemini prompts.
Role modifiers
Open a prompt with a role to set Gemini's perspective, vocabulary, and standards. Put it at the start of the Intent line.
| Role phrase | What it changes | Use for |
|---|---|---|
| Act as a senior copywriter | Punchy, benefit-led prose | Ads, landing pages, emails |
| Act as a staff software engineer | Idiomatic code, edge cases, trade-offs | Code review, architecture |
| Act as a data analyst | Structured findings, caveats on the data | Spreadsheets, reports, charts |
| Act as a strict editor | Tightens, cuts, flags weak lines | Rewrites, proofreading |
| Act as a patient tutor | Step-by-step, checks understanding | Studying, explanations |
| Act as a skeptical reviewer | Pushes back, lists risks and gaps | Plans, proposals, decisions |
| Act as a product manager | User stories, priorities, scope | Specs, roadmaps, tickets |
| Act as a subject-matter expert in [field] | Domain terms, current best practice | Technical or niche topics |
Output-format modifiers
Gemini 3.1 follows explicit format instructions reliably. Add one of these to the Constraints line so the output arrives in the shape you need.
| Say this | You get |
|---|---|
| Return a table with columns X, Y, Z | A clean markdown table, no preamble |
| Answer in bullet points, one idea each | Scannable list, no long paragraphs |
| Return valid JSON matching this schema: {...} | Parseable JSON only, no prose around it |
| Give me numbered step-by-step instructions | Ordered steps you can follow in sequence |
| Keep it under [N] words | A hard length cap Gemini respects |
| Write it as a short paragraph, no lists | Flowing prose instead of bullets |
| Output only the code, no explanation | A raw code block, ready to paste |
| Use headings and a one-line summary at top | A structured document with a TL;DR |
Thinking levels
Gemini 3.1 lets you set how hard it reasons before answering — Low, Medium, or High. Higher levels are slower and cost more, so match the level to the difficulty.
| Level | Best for | Trade-off |
|---|---|---|
| Low | Quick lookups, rewrites, simple summaries, tone changes | Fastest and cheapest; weak on multi-step logic |
| Medium | Most everyday work — drafting, analysis, following multi-step instructions | Balanced default for daily use |
| High | Hard math and logic, code architecture, multi-document synthesis, high-stakes answers | Slowest and priciest; overkill for trivial tasks |
Rule of thumb: if a wrong answer is expensive or the problem has several moving parts, raise to High — ideally on Gemini 3.1 Pro. For volume and speed, keep 3.1 Flash on Low or Medium.
Feature triggers
These phrases switch on specific Gemini features. Include one when you want live data, your own files, a full report, a canvas, or an image.
| Phrase in your prompt | Feature it invokes |
|---|---|
| "Search the web", "as of today", "cite sources with links" | Grounding with Google Search (live results) |
| "Summarize the attached Doc", "@ mention a file", "the numbers in my Budget Sheet" | Workspace files (Gmail, Docs, Sheets, Slides, Drive, Calendar) |
| "Find the email from [name] about [topic]" | Gmail search inside Workspace |
| "Do deep research on", "give me a cited report on" | Deep Research (multi-source, several minutes) |
| "Run the most thorough research possible" | Deep Research Max (widest, longest run) |
| "Open this in Canvas", "let's edit this together" | Canvas (side-by-side editable draft/code) |
| "Generate an image of", "create a diagram / SVG of" | Image generation, native SVG and 3D rendering |
Combine triggers with ICC — for example, an Intent of "give me a cited report" plus Context of "for a US B2B audience" plus Constraints of "under 1,000 words, table of vendors" gives Deep Research a tight brief. See how to prompt Gemini for Deep Research for the full workflow.
Gems — what to save
A Gem is a saved custom assistant that keeps its instructions across every chat. Build one for any task you repeat, and put your standing ICC setup inside it so you never retype it.
| Put this in a Gem | Why |
|---|---|
| Role and expertise ("Act as our brand copywriter") | Locks the perspective for every reply |
| Standing context — audience, product, brand voice | Saves repeating background each time |
| Default constraints — format, length, banned words | Output arrives shaped correctly by default |
| Reference files or examples of good output | Gives the Gem a quality bar to match |
| A short checklist it must run before answering | Enforces consistency (e.g. "always cite sources") |
Good Gems to build: a brand-voice writer, a code reviewer, a study tutor for one subject, a meeting-notes summarizer wired to your Workspace. Use a one-off prompt for anything you only do once.
Example prompts
Six complete prompts assembled from the tables above — a role, a format, and where useful a feature trigger, all wrapped in ICC.
1. Grounded market summary
Act as a market analyst. Give me the current state of the consumer robot-vacuum market. Search the web and use up-to-date sources as of today. Return a table with columns: Brand, Flagship model, Price band, One standout feature. Then add three bullet points on where the market is heading. Cite your sources with links. Keep it under 300 words.Best for: fast, sourced overviews. (Grounding on, Medium thinking)
2. Workspace report summary
Act as a data analyst. Summarize the attached Q2 Performance Report Doc for our leadership team. Context: they have five minutes and care about revenue, churn, and one risk. Constraints: a one-line summary at the top, then three headed sections (Revenue, Churn, Risk), each two sentences max. No jargon. Pull exact numbers from the document.Why it works: names the file, sets the audience, and caps the length. (Workspace on)
3. Deep Research brief
Do deep research on the best project-management tools for a 20-person design agency in 2026. Context: we use macOS, live in Slack, and need Gantt views plus time tracking. Constraints: return a cited report under 1,000 words with a comparison table (Tool, Price per user, Gantt, Time tracking, Verdict), then a single recommendation with reasons. Flag anything you are unsure about.Best for: vendor and tooling decisions. (Deep Research, High thinking)
4. Code review with High thinking
Act as a staff software engineer doing a careful code review. Context: this is a Python function that batches API calls; it will run in production against a rate-limited endpoint. Review it for correctness, edge cases, and rate-limit handling. Constraints: list issues as numbered items, each with severity (high/medium/low) and a one-line fix. Then show the corrected function, code only. [paste your code below]Best for: catching real bugs. (Gemini 3.1 Pro, High thinking)
5. Study explainer as a table
Act as a patient tutor. Explain the difference between mitosis and meiosis to a first-year biology student. Context: I understand basic cell structure but mix these two up. Constraints: return a comparison table with columns: Feature, Mitosis, Meiosis. Then give one plain-English memory trick. Keep the whole answer under 200 words.Why it works: role plus a strict format keeps it simple and scannable. (Low thinking)
6. Brand email in JSON
Act as our senior copywriter. Write a re-engagement email for lapsed free-trial users. Context: friendly, plain-spoken brand voice; the product is a budgeting app; offer is 30% off the first year. Constraints: return valid JSON only, matching {"subject": string, "preheader": string, "body": string}. Body under 120 words, one call to action, no exclamation marks, no hype words.Best for: plugging output straight into an email tool. (Medium thinking)
Want ready-made skeletons instead of writing from scratch? Grab Gemini prompt templates, or browse the best Gemini prompts for finished examples.
Frequently Asked Questions
What is the ICC prompt formula for Gemini?
ICC stands for Intent, Context, Constraints. State what you want (Intent), give Gemini the background and any files it needs (Context), then set the rules for the output — format, length, tone, and what to avoid (Constraints). Writing all three in order removes most of the guesswork that causes vague answers.
When should I raise Gemini's thinking level?
Use Low for quick lookups, rewrites, and simple summaries. Use Medium for most everyday work — drafting, analysis, and multi-step instructions. Raise to High for hard reasoning: math and logic, code architecture, multi-document synthesis, and anything where a wrong answer is expensive. High is slower and costs more, so do not use it for trivial tasks.
How do I force Gemini to use Google Search?
Ask for current information explicitly — phrases like "search the web for", "using up-to-date sources", "as of today", or "cite your sources with links" trigger grounding with Google Search. Gemini 3.1 will then run live searches and attach source links you can verify.
Can Gemini read my Google Drive, Gmail, and Docs?
Yes, when Workspace integration is enabled for your account. Reference files with an @ mention or name them plainly — "summarize the attached Q3 report", "find the email from Priya about the launch", or "pull the numbers from my Budget 2026 Sheet". Gemini reads Gmail, Docs, Sheets, Slides, Drive, and Calendar with your permission.
When should I build a Gem instead of writing a prompt?
Build a Gem when you repeat the same kind of task with the same rules — a brand-voice writer, a code reviewer, a study tutor. A Gem is a saved custom assistant that carries its role, context, and constraints every time, so you skip re-typing the setup. Use a one-off prompt for anything you only do once.
What is Deep Research and when should I use it?
Deep Research is a Gemini mode that browses many sources, plans its search, and returns a structured, cited report over several minutes. Use it for literature reviews, market scans, and vendor comparisons — anything needing breadth across sources. Deep Research Max runs longer and wider for the most demanding reports. For a quick fact, a normal grounded prompt is faster.
Do format modifiers like "return a table" actually work?
Yes. Gemini 3.1 follows explicit output-format instructions reliably. Ask for a table, numbered steps, JSON, bullet points, or a strict word count and it will honor the structure. Put the format instruction in the Constraints part of your prompt so it is not buried in context.
What is the difference between 3.1 Pro and 3.1 Flash?
Gemini 3.1 Pro is the stronger reasoning model for complex, high-stakes work and benefits most from High thinking. Gemini 3.1 Flash is faster and cheaper, tuned for high-volume, latency-sensitive tasks like quick drafts, classification, and short summaries. Both share the 1M-token context window.