Gemini 2.5 Pro Prompts: 40 Copy-Paste Prompts for Research, Coding, and Video (2026)

Gemini 2.5 Pro scores 90% on MMLU and 78% on SWE-bench Verified. Those numbers are competitive, but they don’t explain why the model feels different to use — the 1-million-token context window and native multimodal input do. This library covers 40 prompts across four domains: research and synthesis, code review and generation, image analysis, and video generation via Veo 3.1 Lite. Each prompt includes the exact text and one operator note on why the structure works. If you’re looking for prompts designed for Claude or ChatGPT instead, see our Claude Sonnet 4.6 prompt library or the cross-model prompt pack.

What Makes Prompting Gemini 2.5 Pro Different From GPT-4o or Claude?

Three structural differences change how you write prompts for Gemini 2.5 Pro compared to other frontier models.

1. Context depth, not just context length. Other models with long context windows lose coherence as inputs grow. Gemini 2.5 Pro maintains structured recall across 200K tokens reliably; degradation becomes measurable at 500K+. That makes it the right choice for loading full regulatory filings, complete codebases, or stacked research papers — tasks where GPT-4o would be truncating and Claude Sonnet would start “forgetting” late-document details.

2. Native multimodal input. You can mix text, images, PDFs, and code in a single prompt. This eliminates the OCR-then-analyze pipeline most workflows require. Drop a chart, a spreadsheet screenshot, and a competitor’s website screengrab into one request.

3. Veo 3.1 integration via the API. Text-to-video generation is accessible directly through Gemini API calls. Veo 3.1 Lite is the budget tier at $0.05–$0.08/sec of generated video. Prompts must be structured (subject → action → setting → camera → style) or the model defaults to the most generic interpretation of your input.

10 Research Prompts

These prompts are designed for Gemini Advanced or the Gemini 2.5 Pro API. They exploit the long-context window to process full documents rather than summaries. For our head-to-head comparison of Gemini 2.5 Pro vs Perplexity Sonar Pro on research tasks, Gemini won on citation accuracy but lost on recency for queries about events from the past 30 days.

Prompt R1: Multi-Document Synthesis

I'm uploading three annual reports (text below). Extract all mentions of AI investment across all three. Calculate year-over-year change in AI R&D spend for each company. Produce a 500-word synthesis and a data table with rows = companies, columns = year, AI spend, YoY delta. Flag any discrepancies in how each company categorizes AI vs. general software R&D.

[Paste document text here]

Why it works: Explicit output format (synthesis + table) prevents the model from choosing structure for you. Flagging the categorization ambiguity surfaces the real comparison problem that generic summaries miss.

Prompt R2: Competitor Intelligence

Act as a competitive intelligence analyst. Based on the following 8 press releases from [Company] over the past 12 months (text below), identify: (1) their 3 strategic priorities with supporting evidence, (2) markets they are entering vs. exiting, (3) any signals of M&A activity or talent acquisition. Cite the specific release number for each finding.

[Paste releases here]

Why it works: Requiring citation of the specific release prevents inference beyond the provided text — critical for competitive research that might influence business decisions.

Prompt R3: Citation Accuracy Checker

Here is a research paper (full text below). Identify every factual claim that references a study or statistic. For each: (1) reproduce the claim verbatim, (2) state whether the cited statistic matches what the citation actually says — mark "UNVERIFIABLE" if the source isn't in the provided text, (3) note whether the claim is hedged or stated as established fact. Output as a numbered list.

[Paste paper text here]

Why it works: The “UNVERIFIABLE” instruction stops the model from fabricating plausible-sounding source content — the most common failure mode on citation-checking tasks.

Prompt R4: Market Sizing from Raw Data

Using the CSV data below, estimate total addressable market for [product category] in the US. Show your assumptions as a numbered list before the calculation. Round all figures to 2 significant figures. Where the data doesn't support an assumption, label it "DATA GAP" rather than interpolating.

[Paste CSV here]

Why it works: The “DATA GAP” instruction prevents confident-sounding estimates built on unverifiable assumptions — a problem that undermines TAM calculations when used for board decks or investor materials.

Prompt R5: Technical Explainer for Non-Technical Audience

I'm attaching 2 technical specification PDFs (text below). My audience is a product manager with no engineering background. Write a 600-word explainer covering: what this technology does, who it's for, the 3 most important technical constraints, and 2 practical use cases. Avoid jargon. Define any acronym on first use.

[Paste spec text here]

Why it works: Specifying the audience persona prevents technical hedging that confuses a PM audience. The word count ceiling prevents over-explanation.

Prompt R6: Interview Prep from JD + Earnings Call

I have an interview with [Company] for [Role] in 5 days. Using the job description (JD below) and the company's most recent earnings call transcript (transcript below), generate: (1) 10 questions they are likely to ask with STAR-format answer skeletons, (2) 5 questions I should ask the interviewer that signal I've read the earnings call.

[Paste JD and transcript here]

Why it works: Grounding questions in the actual earnings call forces specificity — questions about real business challenges the company named, not generic “where do you see yourself in 5 years” prep.

Prompt R7: Counter-Argument Generator

Here is my thesis: [paste thesis]. Generate the 5 strongest counter-arguments a rigorous skeptic would make. For each: (a) state the objection in one clear sentence, (b) describe what type of evidence would support the objection, (c) suggest how I can address it without dismissing it. Do not steelman weak objections — only generate objections with real merit.

[Paste thesis here]

Why it works: The “do not steelman weak objections” instruction focuses the model on genuine challenges rather than easy-to-answer strawmen.

Prompt R8: Timeline Constructor

Extract all dates, events, and named actors from the following article. Output as a chronological timeline in this format: [YYYY-MM-DD | Actor | Event | Significance (1 sentence)]. If a date is approximate, flag it as "circa." Do not infer dates not explicitly stated in the text.

[Paste article here]

Why it works: The strict format and “do not infer” rule prevent the model from filling in plausible-but-unverified dates from training data — a common failure on historical event timelines.

Prompt R9: Regulatory Risk Extractor

Here is a regulatory filing (text below). Extract every instance where the company discloses a risk related to AI regulation, data privacy, or model liability. For each: (1) quote the exact language, (2) give the section heading, (3) classify as [FINANCIAL RISK | OPERATIONAL RISK | REPUTATIONAL RISK]. Do not paraphrase — verbatim quotes only.

[Paste filing text here]

Why it works: The “verbatim quotes only” rule prevents the model from softening legal language in ways that could obscure material risk disclosures — critical for compliance or due diligence workflows.

Prompt R10: Literature Review Clustering

Here are 15 academic abstracts on [topic] (below). For each: (1) identify the key methodology in 10 words, (2) state the main finding in one sentence, (3) note any limitations the authors explicitly acknowledged. Then group the 15 abstracts into 3–4 thematic clusters. Write a 100-word introduction for each cluster.

[Paste abstracts here]

Why it works: Processing all 15 abstracts in one call exploits Gemini’s long-context strength. Asking for 10-word methodology summaries prevents over-description that inflates output length without adding analytical value.

10 Coding Prompts

Gemini 2.5 Pro performs at 78% on SWE-bench Verified — below ChatGPT o3 at 88.6% and Claude Sonnet 4.6 at 82.1%, but the 1M token context window makes it the right tool for large-codebase tasks where other models truncate. These prompts are optimized for Gemini Code Assist or direct API access.

Prompt C1: Security-Focused Code Review

Review this Python script for: (1) security vulnerabilities — SQL injection, XSS, insecure dependency versions, hardcoded secrets; (2) performance bottlenecks; (3) readability issues. Output findings in three labeled sections. Each finding must include: severity [HIGH/MEDIUM/LOW], the specific line or function, and a concrete fix. Do not rewrite the full file — flag and explain only.

[Paste code here]

Why it works: Separating security, performance, and readability prevents the model from mixing severity levels and makes the output triageable by different team members.

Prompt C2: Refactor Without Behavior Change

Refactor this JavaScript function for readability. Rules: (1) extract any block over 15 lines into a named helper, (2) replace magic numbers with named constants at the top, (3) remove comments that restate the code, (4) do not change observable behavior. Show a before/after diff for each change — not the full file.

[Paste function here]

Why it works: The “do not change behavior” and “diff only” constraints prevent the model from shipping a clever rewrite that breaks tests — the most common failure mode on refactoring requests.

Prompt C3: Root-Cause Debug

Here is a Python stack trace and the full code file. Identify the root cause — not just the line that threw the error. Explain why it fails. Provide the minimal code change to fix it. Then scan the rest of the file: does the same bug pattern exist anywhere else?

Stack trace: [paste]
Code: [paste]

Why it works: The “scan for the same pattern” instruction catches sibling bugs that the same root cause often produces — skipping this step is why bug fixes frequently require follow-up patches.

Prompt C4: Parametrized Unit Tests

Write pytest unit tests for this function. Requirements: (1) one test for the happy path, (2) parametrized tests for: empty input, null, type mismatch, (3) one test that documents a known failure mode in its docstring. Use pytest.mark.parametrize to reduce duplication. No test should import anything beyond pytest and the module under test.

[Paste function here]

Why it works: The import constraint prevents over-engineered test setups that create maintenance overhead for a future engineer who needs to understand what the function is actually supposed to do.

Prompt C5: Async API Client

Generate a Python module wrapping this REST API (OpenAPI spec below). Requirements: class-based client, async/await using httpx, retry with exponential backoff on 429 and 503, typed response objects using dataclasses, docstring on every public method. Only httpx and stdlib — no additional dependencies.

[Paste OpenAPI spec here]

Why it works: The “stdlib only” constraint produces a client that won’t break when a transitive dependency updates — critical for production code that needs to stay stable.

Prompt C6: SQL Query Optimizer

This query runs in over 8 seconds on a 50M-row table. Analyze the query plan issues. Suggest specific indexes with the exact CREATE INDEX statement. Rewrite the query to use them. Also flag any N+1 patterns if this is part of an ORM. Show that your rewrite returns equivalent results to the original.

[Paste query and schema here]

Why it works: Asking for the CREATE INDEX statement (not just “add an index”) forces specific, actionable output that can be copy-pasted into a migration file.

Prompt C7: Regex Builder with Verification

Write a Python re-module-compatible regex that matches: [precise description]. Handle these edge cases: [list them]. Output: (1) the pattern, (2) a line-by-line breakdown of each component, (3) 5 test strings — 3 that should match, 2 that should not — with expected results and explanation of why.

Why it works: The test strings with expected outcomes make the regex immediately verifiable — regex that “looks right” fails on edge cases constantly, and having the test set inline means you can run it without writing a separate test file.

Prompt C8: Architecture Documentation from Code

Based on these key files (below), generate an architecture document in Markdown: (1) system components and their responsibilities, (2) data flow between components as an ASCII diagram, (3) external dependencies with a one-sentence rationale for each, (4) known tech debt — mark explicitly as [TECH DEBT]. Do not describe implementation details — only the system shape.

[Paste file list and contents here]

Why it works: The “system shape only” instruction prevents the documentation from becoming a code mirror that goes stale as soon as any implementation changes. Architecture docs that describe behavior rather than code last longer.

Prompt C9: Dependency Migration Checklist

I need to migrate from [Old Library] to [New Library] across 45 Python files (imports listed below). Generate: (1) a migration checklist ordered by risk — breaking API changes first, (2) a grep/sed search pattern to find every usage site, (3) the equivalent new API call for each deprecated method I'm using. Note any methods in the old library that have no direct equivalent.

[Paste import list and old library usage here]

Why it works: Ordering by risk means the migration doesn’t fail halfway through when the highest-impact breaking change surfaces too late in the process.

Prompt C10: Static Performance Analysis

Analyze this function's performance without profiling data. Identify: (1) computational complexity of the current implementation using Big-O notation, (2) the two most expensive operations and why, (3) two concrete optimizations with their expected improvement and any correctness tradeoffs. Show the optimized version alongside the original.

[Paste function here]

Why it works: The “without profiling data” framing forces the model to reason about algorithmic complexity rather than guessing hot paths — useful early in development before you have production traces.

10 Image Analysis Prompts

These prompts work in Gemini Advanced (upload via the chat interface) or via the Gemini API with base64-encoded image input. Gemini 2.5 Pro handles charts, screenshots, PDFs, and product images in a single request without separate OCR preprocessing.

Prompt I1: Screenshot to Action Items

I'm sharing a SaaS dashboard screenshot. Extract: (1) all metrics visible with their current values, (2) any alerts or anomalies visually highlighted, (3) 3 specific actions a product manager should take based on what is shown. Format as a prioritized action list with one-sentence rationale per item.

[Attach screenshot]

Why it works: Asking for prioritized actions (not just observations) converts a passive screenshot summary into something a PM can act on in a standup.

Prompt I2: Chart Data Extraction

Extract all numerical data from this chart image as a CSV table with column headers matching the axis and legend labels. Where exact values aren't labeled, prefix your estimate with "~". Note the chart title, axis labels, and any footnotes. Flag any data point where the label is ambiguous or the value is unclear.

[Attach chart image]

Why it works: The “~” prefix for estimates creates a clear distinction between read values and interpolated ones — critical if you’re using the extracted data in a downstream calculation or report.

Prompt I3: UX Audit Against Heuristics

Audit this UI screenshot using Nielsen's 10 usability heuristics. For each violation: (1) name the specific heuristic, (2) describe the exact UI element that violates it, (3) suggest a concrete fix. Focus on the top 5 issues that would most affect task completion for a first-time user. Skip minor aesthetic preferences.

[Attach screenshot]

Why it works: Grounding the audit in named heuristics (not general “UX feedback”) makes findings defensible in design reviews and prevents the model from defaulting to personal aesthetic opinions.

Prompt I4: Document OCR and Summary

This image is a scanned document. Extract all text, preserving structure: headers, bullet lists, numbered sections, tables. Then write a 150-word summary of the key points. Flag any text that is illegible or where you are uncertain of the reading with [UNCLEAR].

[Attach document image]

Why it works: Preserving document structure before summarizing prevents the model from silently dropping tables or numbered lists that may contain the document’s most important content.

Prompt I5: Product Image Copywriting

Based on this product image, write: (1) a 120-character headline for Google Shopping, (2) a 300-word product description for an e-commerce listing emphasizing the 3 most visible features, (3) 5 alt-text variations — each optimized for a different search intent (visual search, accessibility, commercial intent, informational, navigational).

[Attach product image]

Why it works: The 5 alt-text variations surface how different audiences “see” the same product — useful for A/B testing and for SEO teams that need to cover multiple intent clusters from one SKU image.

Prompt I6: Competitor Ad Analysis

This is a screenshot of a competitor's paid ad. Analyze: (1) the primary CTA and the audience segment it's targeting, (2) the unique value proposition being emphasized, (3) any social proof signals visible, (4) what this positioning implies about their current growth strategy. Suggest 2 counter-positioning angles we could test.

[Attach screenshot]

Why it works: Asking for counter-positioning angles (not just ad analysis) turns a passive competitive observation into a brief for the next creative test.

Prompt I7: Whiteboard Transcription

This is a photo of a whiteboard from a strategy session. Transcribe all text verbatim. For diagrams: describe them using structured notation — [Box: Label] → [Box: Label] for flows, (Circle: Label) for decisions. Mark any text you cannot read clearly as [UNCLEAR: your best guess]. Preserve the spatial relationship between elements where visible.

[Attach whiteboard photo]

Why it works: The structured notation convention for diagrams produces output that’s readable in plain text and can be fed directly into a diagramming tool like Mermaid or Lucidchart.

Prompt I8: Cross-Page Brand Consistency Check

I'm uploading screenshots of 4 different pages from the same website. Identify all visual inconsistencies across pages in: (1) typography — font family, size, weight, (2) color usage — note exact inconsistencies, not general palette observations, (3) button styles — shape, padding, color states, (4) spacing — margin/padding patterns. Output as a table: Page | Element | Inconsistency | Recommended Fix.

[Attach 4 screenshots]

Why it works: The table format structures findings so a developer can work through them systematically rather than hunting through prose for action items.

Prompt I9: Logo Brand Analysis

Analyze this logo for: (1) color psychology — what emotions the palette is designed to evoke, (2) shape language — whether the forms are geometric, organic, or angular and what each implies about the brand positioning, (3) typographic personality if text is present, (4) which industry this logo would feel native to, and one industry where it would feel out of place.

[Attach logo image]

Why it works: The “industry where it would feel out of place” question surfaces the implicit brand territory the logo is occupying — useful when evaluating whether a visual identity signals the right market to a target audience.

Prompt I10: WCAG Accessibility Pre-Check

Review this UI screenshot for accessibility issues. Identify: (1) any text that likely fails WCAG 2.1 AA contrast ratio (4.5:1 minimum for normal text), (2) interactive elements that appear smaller than 44×44px minimum touch target, (3) any information conveyed only through color with no secondary indicator, (4) any interactive elements that appear to lack a visible focus state. Output as a checklist with pass/fail/uncertain for each criterion.

[Attach screenshot]

Why it works: Pass/fail/uncertain output gives developers a triage priority — uncertain items go to automated WCAG tooling for verification rather than wasting review time on elements that clearly pass.

10 Video Prompts for Veo 3.1

Veo 3.1 prompts follow a five-part structure: Subject + Action + Setting + Camera Movement + Visual Style. The Lite tier ($0.05–$0.08/sec) handles this structure well. Keep prompts under 200 words — longer inputs cause Veo 3.1 to weight later instructions more heavily, often overriding the camera and style specifications you opened with.

Prompt V1: Product Explainer

A minimalist desk with a smartphone displaying a clean SaaS analytics dashboard. The phone screen activates, showing rising metrics. A hand enters frame and swipes through chart screens. Modern home office, natural daylight from left. Slow push-in camera movement. Corporate-professional visual style, warm but neutral color grade. 10 seconds.

Why it works: “Slow push-in” is the camera movement that communicates credibility without motion sickness — avoid handheld or zoom-in for product demos where you want the UI to be readable.

Prompt V2: Developer Tutorial Aesthetic

Close-up of hands typing on a mechanical keyboard. Dark-mode code editor fills the background screen. A terminal window opens; green deployment success text scrolls upward. Shallow depth of field, focus rack from keyboard to screen midway. Slightly warm cinematic color grade. No face visible. 12 seconds.

Why it works: “No face visible” avoids deepfake-adjacent content flags that some platforms apply to AI-generated human faces in developer content.

Prompt V3: Social Media Short (15s)

Split-screen vertical format. Left side: person overwhelmed at a cluttered desk, papers everywhere, frustrated expression. Right side: same person relaxed at a clean desk, using a laptop, smiling. Text overlay fades in at 5 seconds: "Before AI. After AI." Bright, high-saturation color grade. Quick-cut rhythm, cuts at 7.5 seconds to switch scenes. 15 seconds.

Why it works: Specifying the text overlay timing (5 seconds in) gives the scene time to establish before the message appears — the most common mistake in short-form video is overlaying text too early.

Prompt V4: B-Roll Lifestyle Footage

Aerial shot of a city skyline at dusk, city lights beginning to appear, buildings catching the last of the sunset. Camera slowly tilts down from skyline to street level over 8 seconds. Golden hour lighting. No individual people in focus — crowd-level anonymity. Cinematic color grade, slightly desaturated with warm shadow tones. 10 seconds.

Why it works: “Crowd-level anonymity” avoids identifiable-face generation, which can complicate rights clearance for commercial use.

Prompt V5: UI Screen Recording Style

Screen-capture aesthetic: a cursor navigates a clean web application — clicks through a dashboard menu, selects a filter, table data updates instantly. Subtle motion blur on fast cursor movements. Clean neutral desktop background visible at screen edges. Enterprise software aesthetic — no gradients, flat UI. No voiceover. 12 seconds.

Why it works: “Screen-capture aesthetic” signals a motion style that Veo 3.1 interprets correctly — without this framing, the model defaults to filming a screen from outside rather than simulating the UI in motion.

Prompt V6: Interview / Testimonial Format

Person in their early 30s, professional dress, seated in a bright modern open-plan office. Speaking directly to camera, relaxed posture, occasional natural hand gestures. Shallow depth of field, background workstations subtly blurred. Warm natural light from camera-left window. Documentary interview style. No text or graphics. 10 seconds.

Why it works: “Relaxed posture, occasional natural hand gestures” produces a more authentic performance than prompting for “confident” or “professional” — behavioral descriptors outperform personality adjectives in Veo prompts.

Prompt V7: Animated Data Story

Animated bar chart: 5 bars grow from baseline upward in sequence, left to right, each labeled below with a category name. Numbers count up at the top of each bar as it grows. Color-coded — 3 distinct hues from the same palette family. Clean white background. Modern sans-serif typography throughout. Data journalism visual style. 12 seconds.

Why it works: “Data journalism visual style” is a style reference Veo 3.1 interprets consistently — it produces the NYT/FT graphics aesthetic without requiring you to describe every design element individually.

Prompt V8: Conference Highlights Opener

Quick montage in this sequence: (1) empty conference hall, chairs in rows, stage lit, (2) attendees filing in, time-lapse effect, (3) two professionals networking, smiling, (4) speaker at podium, wide shot, (5) close-up of attentive audience member taking notes. Energetic cross-cut rhythm, cut every 2 seconds. Warm high-contrast color grade. No text overlays. 10 seconds.

Why it works: Specifying the sequence as a numbered list gives Veo 3.1 a clear shot order to follow — without explicit sequencing, the model generates a random montage order that may not tell the intended story.

Prompt V9: Brand Logo Reveal

Logo reveal animation: starts with abstract light particles drifting across a dark background. Particles converge and slow, forming a geometric shape that morphs into a company logomark. The brand name fades in beneath it as particles dissolve. Clean dark background with a subtle blue-to-dark gradient. Premium minimalist aesthetic. No sound described — audio will be added separately. 6 seconds.

Why it works: “No sound described” keeps the Veo output as a clean visual that can be matched to custom audio in post — Veo 3.1 audio sync is inconsistent at 6-second durations, so separating audio production gives better final results.

Prompt V10: Café Ambient with Audio (Veo 3.1 Audio Mode)

Interior of an independent café, morning. Warm light through large front windows. A barista steams milk at a semi-automatic espresso machine — steam wand sound, the knock of a portafilter. Ambient chatter in the background. Camera slowly pans from the espresso counter to a window table where a laptop is open. Documentary realism, warm tones. Audio: café ambience — espresso machine, milk steaming, low background conversation. 15 seconds.

Why it works: Veo 3.1 audio mode requires the audio scene description to be explicit (not implied by the visual scene). Listing specific sounds — “espresso machine, milk steaming, low background conversation” — produces a more accurate audio mix than generic “café sounds.”

Common Prompting Mistakes That Cost Quality

Three failure patterns appear consistently when teams move from GPT-4o workflows to Gemini 2.5 Pro without adjusting their prompts.

Summarizing before you send. Teams pre-summarize documents to stay “within context limits” — a habit from working with GPT-4o (128K context) that’s counterproductive with Gemini 2.5 Pro’s 1M token window. Summaries lose structure, remove tables and figures, and introduce the original author’s interpretation into your analysis. Send the raw source.

Asking for “thorough” responses. Gemini 2.5 Pro interprets “thorough” as permission to pad. Specify length explicitly: “Write a 400-word response,” or “Use 3 paragraphs maximum.” The model responds well to hard word count ceilings.

Skipping format specification on structured outputs. When the output needs to be JSON, CSV, or a Markdown table for downstream processing, specify the format in the first sentence of the prompt — not the last. Gemini begins generating before reading the full prompt in streaming mode, and a late format instruction arrives after the model has already committed to prose structure.

Per Google DeepMind’s Gemini 2.5 Pro technical overview, the model is designed for “complex reasoning, coding, and multimodal tasks” within a 1,048,576-token context window — the widest context available among frontier models as of Q2 2026, enabling full-document analysis without preprocessing or chunking for most enterprise document types.

Frequently Asked Questions

Last updated: 2026-07-31