Wrangl Docs
AI agents

Write a prompt

Choose a model and a mode, use slash commands and mentions of files and other chats, attach files, and watch how much context is left.

The message box at the bottom of a chat is where you talk to the agent. Besides text, it holds the controls for the model, how much the agent may do on its own, and what you send along with your words. The launcher uses the same box, so everything on this page works there too.

Send and add lines

  • Press Enter to send. Ctrl+Enter (Cmd+Enter on macOS) also sends.
  • Press Shift+Enter to start a new line.

If the agent is still working when you send, Wrangl queues your message and sends it when the agent finishes. See Queue prompts.

Change the model, effort, and fast mode

Click the model button. It sits to the left of the mode button and shows the current model, followed by the effort level when the agent has one. The model panel opens with the cursor in its search field.

  • Click a model in the list to switch to it. A check mark shows the current model.
  • Type in the search field to filter models and agents by name. The number of matches shows at the end of the field. Press Enter to pick the first match. When nothing matches, Enter does nothing.
  • Use the up and down arrow keys to move through the list, and Enter to pick a model.

The current model's row has a button on its right. It shows the effort level, or Edit when the agent has no effort levels. Click it, or press the right arrow key on the row, to open the model's options beside the list:

  • Thinking sets how hard the model thinks before it answers. Higher levels are slower and give more careful answers.
  • Context sets the size of the context window, when the agent offers more than one.
  • Fast mode turns fast mode on or off.

Each option appears only when the agent offers it. Press the left arrow key or Escape to close the options and go back to the row. The agent's other settings appear at the bottom of the panel. To hand the chat to a different agent instead, see Switch agents mid-chat.

Keep favorite models at the top

Hover over a model and click its star to add it to Favorites. Favorites appear at the top of the list and can come from different agents. Click the star again to remove a model.

Choose how much the agent may do

The mode button sits between the model button and the send arrow. A mode controls whether the agent asks you before it edits files or runs commands. Each agent has its own list of modes. Wrangl sorts them into groups so you can see how much each one allows:

  • Ask first. The agent asks before it edits files or runs commands, or it only reads and plans. Pick these when you want to check each step.
  • Let agent act, where agent is the agent's name. The agent edits files, and sometimes runs commands, without asking first. Pick these for routine work you will review afterwards.
  • Use with care. The agent does everything without asking, including commands that can delete files or push code. These modes sit at the bottom of the menu, under a line. They have a yellow icon, and their description ends with "Use with care." While one is on, the icon on the mode button is yellow too.

Modes that Wrangl does not recognize are listed under Other modes. Read their descriptions in the menu. Wrangl shows the group names only when the agent's modes fall into more than one group.

The mode menu. Default and Plan are under Ask first. Accept edits is under "Let ... act" with the agent's name. Bypass permissions is at the bottom, below a line, with a yellow icon and the words "Use with care."

The mode button appears only when the agent offers modes.

Run a slash command or a skill

Type / at the start of the box, or after a space, to open the slash menu. Keep typing to filter it. The menu has these sections:

  • Skills. Instruction packages installed on your computer or in the project. See Set up agents to add them.
  • Commands. The commands the agent offers, such as a review command.
  • One section for each agent setting, such as Model, Thinking Level, or Mode. Picking a value here changes the setting right away.

The slash menu above the message box after typing "/rel". It lists the release-notes skill under Skills.

Use the arrow keys to move and Enter or Tab to pick. A skill or a command becomes a pill in the box. Type the rest of your message after it. A message can hold one command. To remove a pill, press Backspace right after it.

Pressing Esc while the slash menu is open closes the menu and also clears everything in the box.

Mention a file or another chat

Type @ to open the mention menu. A bare @ shows two entries:

  • Conversations lists the other chats in this project, newest first. A chat that is still working shows a spinner.
  • Explorer lists the files and folders in the chat's folder, with the files you mentioned Recently at the top.

Pick an entry to open its panel. In Explorer, pick a folder to look inside it, or pick Use this folder to mention the folder itself. To go back, press the left arrow key, press Backspace while the search is empty, or click the panel's title at the top of the menu (it has a back arrow).

To search instead, type part of a name after @. The menu lists matching Files and Folders, plus any Conversations whose title matches.

The file menu after typing "@cart" shows src/cart.ts under Files.

Pick a file, a folder, or a chat to add it as a pill. When you send:

  • For a file or a folder, the agent gets its path, so it can open it itself.
  • For a chat, the agent gets the latest turns of that chat as context. Very long chats are cut to their most recent part. If the chats you mention add up to more than 30,000 characters, the ones that do not fit are sent without their text, and the agent gets a note that each was left out. Your own chat history keeps only the pill, not the copied text.

Press Esc to close the menu and keep what you typed.

Attach files and images

You can attach text files, code files, and images. To attach a file, use one of these:

  • Click Attach files (the paperclip) and choose files.
  • Drag files onto the box. The box shows "Drop files to attach".
  • Paste an image or a screenshot.

Each file shows as a chip above your text. Hover an image chip to see a preview. Click the X on a chip to remove it.

The message box with an attached screenshot named checkout.png and a question about it.

Other file types, such as PDFs, are not supported. Wrangl shows "Unsupported file type (text or image only)".

Files that you drag in or paste, and every file you attach in a browser, go into the message itself. These limits apply to them:

  • A text or code file can be up to 512 KB. The agent must accept files in a message. If it does not, Wrangl asks you to use the paperclip instead, in the desktop app.
  • An image can be up to 10 MB when the agent accepts images.
  • In a browser, the agent must accept images. If it does not, Wrangl shows "This agent can't accept images in the browser".
  • In the desktop app, an image for an agent that does not accept images is saved to a temporary file. The agent gets its path.

A file over the limit shows "File too large".

In the desktop app, text and code files that you pick with the paperclip are sent as paths. The agent reads them from disk. Images you pick go into the message when the agent accepts images, and are sent as paths when it does not.

Check how much context is left

An agent can hold only a certain amount of conversation in its memory, called the context window. When it fills up, the agent forgets or summarizes earlier parts of the chat.

Many agents report how full their context window is. After your first message, a small ring appears in the top-right corner of the message box. The ring fills as the conversation grows. Click it to see the numbers:

  • How much of the window this conversation uses, as a percentage and in tokens.
  • How many tokens remain.
  • The total in context, including the agent's own instructions.
  • The cost, when the agent reports it.

The Context window popover shows 29% conversation used, 130K tokens remaining, and a reported cost.

The ring appears only when the agent reports its usage, and only after the conversation uses about 1% of the window. The model panel shows the same numbers in its Context row.

On this page