Nano GPT Manual
What It Does
Shows the same movable overlay UI while sending prompts, screenshots, and selected server options to Connect-Proxy-1.
Hotkeys
- Right Option: send screenshot to AI
- Right Shift: toggle overlay hide/show
- Left Shift + 1 (!): enter typing mode after clicking Waiting
- Click Waiting first, then Left Shift + 2 (@): enter command mode
- Hold Command: enter selection mode with crosshair
- While holding Command, drag mouse to select region
- While still holding Command, press Right Option to ask AI about selected region
Selection Mode
- Selection rectangle must be at least 5px by 5px
- Start and end markers are 1px
- If Command is released without asking, selection disappears
Typing Mode
- Click Waiting first, then press Left Shift + 1 (!) to start typing mode
- Overlay shows Typing only when input is empty
- After you type content, overlay shows only your typed text
- While in typing mode, hold Command and drag-select to choose a region, then press Enter to ask using typed text on that region
- Press Enter to submit typed question and switch to Answering
- Press Right Option in typing mode to discard current typed text and send full screenshot with default prompt
- Press Esc or left click to cancel typing mode
- Press Right Shift in typing mode to exit typing mode and hide overlay
Command Mode
- Click Waiting first, then press Left Shift + 2 (@) to open command mode
- Placeholder is › Typing, then it becomes › your command as you type
- Press Enter to apply command silently
- Commands never require a leading slash; suggestions work from the plain command name
- Invalid command does nothing
- Press Esc, Right Option, or left click to cancel command mode
- Press Right Shift in command mode to exit and toggle overlay
- Commands are entered locally but parsed on the server: color #RRGGBB alpha, size 1-80, ocr, ans, answer, agent, keepcontext, pausehistory, clearhistory, trueclick, fullpage, profile <fastest|standard|smartest>, effort <off|low|medium|high>
- Connection is fixed to Connect-Proxy-1. Use profile fastest, profile standard, or profile smartest.
- Thinking levels: off, low, medium, high.
Command Autocomplete
- While typing a command, a half-opacity ghost suffix shows the active suggestion
- Press Tab to accept the active suggestion
- Press Arrow Down / Arrow Up to cycle through suggestions
- For fixed-set fields (model id, effort level, top-level keyword) suggestions list every match
- For free-form fields (hex color, size number) one canned example is shown (e.g. color ffffff 1, size 10)
Overlay Behavior
- Default text is Waiting
- Overlay can be dragged and position is saved
- Double-click overlay text to copy answer
- During extension auto-update, overlay shows Updating and then restores previous text
- When extension mode is OFF, overlay is hidden
- When extension mode is ON, overlay is shown
Popup Settings
- Current: Extention ON/OFF with Toggle button
- Current: Answer Mode/OCR Mode with Toggle button
- Connect selector: Connect-Proxy-1. Profile selector: Fastest, Standard, Smartest. Thinking selector: Off, Low, Medium, High.
- Effort selector (dropdown): Default, None, Minimal, Low, Medium, High, XHigh. Levels the current model does not support are grayed out and marked (n/a).
- RGBA inputs (alpha supports decimals)
- 12x12 preview box shows text color and transparency
- Size slider controls overlay font size
- Additional system prompt textarea is persistent and autosaved
- Additional system prompt is appended after the built-in system prompt
- Additional system prompt textarea is vertical-resize only with min/max height bounded to viewport
- Reset position button restores default position
- Open Manual button opens this page
- All settings autosave in storage.sync
AI Output Rule
- System behavior asks for answer only
- Subscript/superscript rendering supports _ and ^ notation, for example H_2O, x^2, x^{10}, a_{ij}
- If no question is found, output is No Question found
中文说明
功能说明
保持相同的可拖动覆盖层界面,并将提示、截图和服务器选项发送到 Connect-Proxy-1。
快捷键
- 右 Option:发送截图给 AI
- 右 Shift:切换覆盖层显示/隐藏
- 左 Shift + 1(!):点击 Waiting 后进入输入模式
- 先点击 Waiting,再按左 Shift + 2(@):进入命令模式
- 按住 Command:进入选区模式(鼠标变十字)
- 按住 Command 时拖动鼠标:框选区域
- 按住 Command 不放,再按右 Option:只针对框选区域提问
选区模式
- 选区最小尺寸为 5px x 5px
- 起点和终点标记为 1px
- 如果松开 Command 且未提问,选框会消失
输入模式
- 先点击 Waiting,再按左 Shift + 1(!)进入输入模式
- 输入为空时覆盖层显示 Typing
- 输入非空后只显示你输入的文字
- 输入模式中可按住 Command 拖拽框选区域,再按 Enter 以“输入问题 + 选区”提交
- 按 Enter 提交输入问题并切换到 Answering
- 输入模式下按右 Option 会丢弃当前输入,并按默认提示发送整屏截图
- 按 Esc 或左键点击可取消输入模式
- 输入模式下按右 Shift 会退出输入模式并隐藏覆盖层
命令模式
- 先点击 Waiting,再按左 Shift + 2(@)进入命令模式
- 空输入时显示 › Typing,输入后显示 › 你输入的命令
- 按 Enter 静默应用命令
- 命令无效时不做任何改动
- 按 Esc、右 Option 或左键点击可取消命令模式
- 命令模式下按右 Shift 会退出并切换覆盖层开关
- 命令:color #RRGGBB alpha、color RRGGBB alpha、size 1-80、ocr、ans、answer、profile <fastest|standard|smartest>、effort <级别>
- 连接固定为 Connect-Proxy-1;可使用 profile fastest、profile standard、profile smartest。
- effort 级别:none、minimal、low、medium、high、xhigh(或 default 以清除)。当前模型不支持的级别会被静默忽略。
命令自动补全
- 输入命令时会以半透明显示当前建议作为后缀
- 按 Tab 采用当前建议
- 按下方向键 / 上方向键 循环切换建议
- 对于固定取值字段(模型 id、effort 级别、顶层关键字)会列出所有匹配项
- 对于自由输入字段(颜色十六进制、字号数字)只显示一个示例(如 color ffffff 1、size 10)
覆盖层行为
- 默认显示文本为 Waiting
- 覆盖层可拖动,位置会保存
- 双击覆盖层文本可复制当前答案
- 扩展自动更新时覆盖层会显示 Updating,完成后恢复到之前文本
- 模式为 OFF 时覆盖层隐藏
- 模式为 ON 时覆盖层显示
弹窗设置
- Current: Extention ON/OFF + Toggle 按钮
- Current: Answer Mode/OCR Mode + Toggle 按钮
- Connect 固定为 Connect-Proxy-1;Profile 为 Fastest / Standard / Smartest;Thinking 为 Off / Low / Medium / High。
- Effort 选择器(下拉):Default、None、Minimal、Low、Medium、High、XHigh。当前模型不支持的级别会以灰色(n/a)显示。
- RGBA 设置(Alpha 支持小数)
- 12x12 预览框显示当前文本颜色与透明度
- Size 滑块控制字体大小
- 附加 system prompt 文本框可持久化并自动保存
- 附加 system prompt 会追加在内置 system prompt 之后
- 附加 system prompt 文本框仅允许垂直调整,并有最小/最大高度(不超出视窗)
- Reset position 可恢复默认位置
- Open Manual 打开本手册页面
- 所有设置会自动保存到 storage.sync
AI 输出规则
- 系统提示要求只输出答案
- 支持使用 _ 和 ^ 记号渲染下标/上标,例如 H_2O、x^2、x^{10}、a_{ij}
- 如果没有发现问题,输出 No Question found
Agent mode
Switch Response Mode to Agent. Agent mode requires a typed prompt; it will not run from an empty capture. Along with the screenshot, Nano GPT supplies a bounded collapsed HTML outline containing numbered nodes. Form values, password contents, data-* attributes, cookies, storage and hidden page state are not included. The model can request expandHTML for one relevant branch at a time, which keeps token use bounded even on very large DOMs. runJS is computation-only and executes in a disposable Worker inside an invisible extension sandbox with no website DOM, no extension APIs, and network/storage/worker-spawning APIs disabled. The website cannot inspect the code being run, and each worker is forcibly terminated after 1 second. Normal click uses DOM hit-testing at the requested viewport point without moving the physical mouse pointer. True Clicking is off by default and uses trusted browser input when enabled in Settings or with trueclick on. All server-issued commands execute automatically through the fixed packaged client primitives. runJS remains isolated from websites and is killed after 1 second.
Add --keep-context to a prompt to retain consecutive Agent conversations. The retained streak is cleared by the next Agent prompt that does not keep context. Use keepcontext or keepcontext on|off to toggle the default. Add --full-page to force a full-page screenshot for that Agent turn.
History
Use Show history in the popup to open the local history page. Nano GPT keeps up to the most recent 1000 Answer, OCR, and Agent chats. The history page shows full agent tool commands, preview-truncated tool output with a reveal control, and a Run again button for isolated RunJS entries. It can also pause/resume storage and clear all stored history. You can also use pausehistory to toggle history storage or clearhistory to clear it from the command overlay.