Cropped screenshot? Add offsets before scaling

Use Gemini web to check screenshot-to-test-viewport coordinates with formulas and bounds. This is a paper exercise, not control of a real website. Prompts are untested; results require human review.

Execution environment and usage

AI / software

Gemini

Where to operate

Ordinary text chat in the Gemini consumer web app at gemini.google.com. This card uses manually supplied text only.

Getting started

  1. Prefer an adult personal Google account and check regional/account eligibility first. Do not buy a plan for this exercise.
  2. Paste the fictional fixture into ordinary chat, or substitute authorized de-identified short text. Connect no external services.
  3. Compare output with the human acceptance checks. Preserve errors, gaps and unknowns for review.

Setup and plan requirements

Official web documentation checked 2026-11-12 lists Hong Kong, Macau and Taiwan, plus Simplified, Traditional and Hong Kong Chinese and English. This exercise recommends an adult personal Google account. Child, school and work accounts require separate eligibility, administrator and plan checks; availability is not assured for every account.

Prompt portability limits

Start with standard free access; no subscription is required for this card. Taiwan advertises NT$0/month for free access. Local Hong Kong/Macau paid prices are unverified, so no local paid price is promised. Compute limits vary with model, complexity and conversation length, including five-hour and weekly caps; no fixed number of prompts is guaranteed. Wait if limited rather than assuming an upgrade is needed. Mobile, API, CLI and the original automation have separate eligibility, account and cost requirements.

Usage limitations

Documentation review is not a test from the three regions. Prompts, account access, model output and the original full workflow were not tested. Use fictional or de-identified data, never confidential inputs; some data can receive human review. Disabling activity is not zero retention: relevant chats can remain for 72 hours and undergo safety review. A person must verify outputs; text advice is not execution.

Environment sources

適用情境 / When to use

Use Gemini web to check screenshot-to-test-viewport coordinates with formulas and bounds. This is a paper exercise, not control of a real website. Prompts are untested; results require human review. Fictional fixture, not a real usage record: Full screenshot 1600×1000 pixels maps to an 800×500 CSS viewport. Crop starts at full-image (200,100), size 800×600. Target center is crop-local (400,300). Assume no scrolling, browser chrome or additional distortion.

步驟 / Steps

  1. Preserve input conditions and existing labels and use the fictional fixture with the parameterized prompt in text chat.
  2. Human check: full-image point (600,400); both scales 0.5; CSS point (300,200). Ignoring crop offsets wrongly gives (200,150). If the point changes to crop-local (900,300), it is outside width 800: reject rather than clamp.
  3. If output fails a check, record the discrepancy and revise the request for human review. Never present an expected answer as observed model output.

提示詞 / Prompt

Given [full-image size], [CSS viewport size], [crop origin and size] and [crop-local point], validate bounds, add crop offsets, then scale each axis. Show units, assumptions and result. Stop guessing if offset, coordinate frame or scale is missing. Execute no clicks.

Fictional fixture, not a real usage record: Full screenshot 1600×1000 pixels maps to an 800×500 CSS viewport. Crop starts at full-image (200,100), size 800×600. Target center is crop-local (400,300). Assume no scrolling, browser chrome or additional distortion.

Also test the same fixture with crop-local point (900,300); report whether it is out of bounds and whether calculation should proceed.

留意事項 / Cautions

  • Platform and region: ordinary Gemini web text chat. Official lists include Hong Kong, Macau, Taiwan, Chinese variants and English. Documentation checked 2026-11-12; local connections were not tested.
  • Plan and account: prefer an adult personal account. Standard free access is a starting point, with variable usage caps and no required subscription. Taiwan advertises NT$0/month free access; Hong Kong/Macau paid prices are unverified. No particular model or prompt count is guaranteed. Work, school and child accounts need separate checks.
  • No original toolchain is connected. Web availability does not establish mobile, API, CLI or original-automation access. Never supply confidential data. Activity-off chats can still remain for 72 hours and receive safety review.
  • The formula holds only under the stated assumptions. It proves neither browser-automation availability nor safety. Do not use it for payments, deletions or unauthorized clicks.
  • The original uses normalized model coordinates; this card instead uses explicit screenshot pixels and CSS coordinates. Do not mix these coordinate frames.

來源限制 / Evidence limits

The original source provides inspiration, not validation of this exercise. Its toolchain, regional services and plans have not been reproduced. This card uses only the documented Gemini web text environment. The source author explains Java browser control and scaling normalized model coordinates to a viewport. Crop offsets and image-pixel/CSS examples are editorial additions, not the source’s normalized 1000-grid. The visible publication date is November 2 despite the November 3 URL. Original publication date: 2025-11-02; checked: 2026-11-12.

流程說明 / Workflow note

Original editorial adaptation, not a newly discovered use case or the source author’s verbatim workflow. Source-reported outcomes do not guarantee this exercise. Prompts were not tested. Adapted from this existing case: https://hkfrees.com/ai/case/reported-browser-coordinate-calibration/?lang=en Use Gemini web to check screenshot-to-test-viewport coordinates with formulas and bounds. This is a paper exercise, not control of a real website. Open this card’s title for the official environment-source links and human acceptance steps on its full case page.

原始工具組合 / Original tool stack

Gemini Computer Use; Java; Playwright

Source publication date: 2025-11-02

Source checked: 2026-11-12

Driving a web browser with Gemini’s Computer Use model in Java