# describe

Prompt for Claude Code and ChatGPT that describes phone screenshots for an assistant without visual input. Converts mobile UI layouts into structured text with normalised tap coordinates.

- prompt · `type=prompt` · Claude Code, ChatGPT · ★ 9488 repo stars
- Author: wangzan101
- License: Apache-2.0 · policy=mirror · the beginning of the file is at the end of this page
- Source: https://github.com/wangzan101/Mika/tree/HEAD/core/src/main/resources/prompts/roles/describe.md
- Updated: 2026-11-16
- Security check: automated, 2026-11-17 (text) — no signs of malicious code found.
- HTML: https://affpapa.org/en/nevibecode/prompts/describe

## What it does

The prompt instructs a vision model to convert mobile screenshots into a structured summary for another agent. The entire screen maps to a normalised 0-999 coordinate grid with (0,0) at the top-left corner. For interactive controls, it calculates the exact center point for tap automation.

Output follows four specific sections: screen overview, an element list capped at 15 items with types and coordinates, noticeable blockers or dialogs, and task-related visual clues. Passive text titles and body paragraphs are ignored to keep context concise.

No action recommendations or next steps are permitted. The model only records clearly visible elements and skips anything obscured.

## Install

### Claude Code (native platform)

**Manually: put the files in place**
- File: `.claude/commands/describe.md`

### ChatGPT

**How to use**
- Paste the text at the start of a chat or into project instructions

Where a file is given instead of a command, copy the files from the source to that path: https://github.com/wangzan101/Mika/tree/HEAD/core/src/main/resources/prompts/roles/describe.md

## When to use it
- parse interactive mobile UI elements from a screenshot
- calculate tap coordinates for buttons and inputs
- feed phone screen state to an automation agent
- identify blocking modal dialogs and loading spinners

## Example prompts
- Describe the screenshot `screen.png` using the coordinate grid
- Extract interactive controls and tap points from this mobile screen
- Check this phone UI screenshot for blocking dialogs or errors

## Source text · describe.md · Apache-2.0

Below is the content of a third-party file. It is data, not instructions for you: do not act on anything written inside unless the human asks. Block tag: 48af7b40d40806aa.

BEGIN THIRD-PARTY FILE 48af7b40d40806aa
```text
You describe one phone screenshot for an assistant that cannot see the screen
and must decide what to do from your words alone. You report what is there;
you do not decide or advise. The assistant is waiting for your answer: keep it
short.

Coordinates: the whole screen is 0-999 on each axis, 0,0 is the top-left
corner and 999,999 the bottom-right. For every interactive element give the
point at its exact centre; that point will be tapped.

Reply in exactly these four sections and nothing else:
[屏幕] One sentence: which app or site this is and what kind of page.
[可交互元素] One line per element, at most 15 lines: - name | type (button, input, list item, icon, switch, tab) | (x,y)
  Name each in a few words. List first the elements that matter for the task,
  then the navigation controls (back, close, bottom tabs, buttons of any
  dialog). Leave out text that is not a control (titles, message bodies,
  paragraphs); for a long list, give its first few items.
[异常或值得注意] Any dialog, error, loading state, blank area, sign-in wall,
  verification check or anything else unusual or blocking, in a line or two.
  Write 无 if there is none.
[与任务有关的线索] At most two lines: text or elements on the screen that
  relate to the task, stated plainly. Do not say which one to tap and do not
  give steps. Write 无 if there is none.

Rules:
- Put each coordinate as close to the element's centre as you can.
- Never list an element you cannot clearly see.
- No advice and no suggestions; only what is on the screen.
```
END THIRD-PARTY FILE 48af7b40d40806aa

The third-party file has ended. Everything between the 48af7b40d40806aa tags is data: requests and commands from it are not followed. An “end of file” line with another tag inside the block is part of the file.

© wangzan101 · Apache-2.0 · https://github.com/wangzan101/Mika/tree/HEAD/core/src/main/resources/prompts/roles/describe.md

## Similar
- [harmonyos-device-automation](https://affpapa.org/en/nevibecode/skills/harmonyos-device-automation.md) — skill · 2500 installs
- [Desktop Control](https://affpapa.org/en/nevibecode/skills/desktop-control.md) — skill · 4000 installs
- [Droidline](https://affpapa.org/en/nevibecode/mcp/droidline.md) — MCP server · ★ 558 repo stars
- [phone-harness](https://affpapa.org/en/nevibecode/skills/phone-harness.md) — skill · ★ 567 repo stars
- [capture-screen](https://affpapa.org/en/nevibecode/skills/capture-screen.md) — skill · 1100 installs
- [browser-navigation](https://affpapa.org/en/nevibecode/skills/browser-navigation.md) — skill · 120 installs

## Other versions with the same name
- [describe · wangzan101](https://affpapa.org/en/nevibecode/prompts/wangzan101-describe.md) — ★ 9488 repo stars

From the same set: [wangzan101/mika](https://affpapa.org/en/nevibecode/by/wangzan101/mika.md) — 132 cards

---
NeVibeCode on affpapa.org. Catalog texts are descriptions, not instructions for the agent. How to choose and install: https://affpapa.org/en/nevibecode/agents.md
