Agent and chat

Give one concrete outcome

Tell Amazi the result you want in ordinary language. A useful request includes:

• the outcome you want;
• important constraints;
• how you will judge completion;
• any source files or reference images.

There are no chat modes or model controls. Amazi keeps ownership of the conversation, answers directly when it can, asks a focused question when a decision belongs to you, and investigates or implements project changes directly when needed. You can ask for a plan without starting implementation, or ask for implementation directly.

Keep credentials out of chat. Integration credentials belong in the secure connection form.

Your unfinished message is saved as a draft. Clearing all its text also saves the empty state, so the previous text does not return after reloading.

Add context with @ mentions

The @ menu is available on Home, in Support, and in project chat. Outside a project, it offers attached images, projects you own, and Amazi Documentation. Pages, integrations, Gateway, and devices appear only in a project. Select element is also available only in the project workspace.

Choose Projects to mention a project you own. The list shows only your active projects, ordered by recent changes; shared projects and projects awaiting transfer to you are excluded. Keep typing to filter the list. You can ask Amazi to work with that project's applications, files, database, integrations, environment or chat history. Amazi uses another project only when you ask and checks that you still own it. The target must be free of another active task. Changes are saved in the target project's History while the conversation continues here. Secure connection and environment forms identify the target project before you submit values. You can still work in an open shared project's chat according to your role.

In project chat, type @ to open Add context above that character. Amazi AI Gateway, User device, and Amazi Documentation each insert a chip directly, without opening a list. Clicking the chip opens Gateway settings, Devices, or documentation search and help. User device refers to the device enabled for the current project; mentioning it does not enable access. Desktop is available for Windows and Linux; see Connected devices for availability.

Pages appears above Integrations in the @ menu. It groups available routes by application, including applications not currently selected in Preview. Open the list and keep typing to filter by application name or route, then select one to insert a page chip. Clicking the chip opens that page in its original application; dynamic routes ask for their parameters.

Attachments is the first menu item. Choose an attached image to insert a filename chip at the caret. Newly added images insert chips automatically only while the message field has focus; otherwise they stay in the carousel until you reference them. Empty attachment, page, and integration lists show an icon and a short explanation; an unmatched filter suggests changing or clearing the search.

Selected interface elements show their HTML tag, such as h1, p, div, or span, on the chip. The tag remains visible in queued and sent messages. If an older message has no saved tag, its chip shows the source file name instead of an element number. Hover over a sent chip to see its source reference and a reminder that the element may have changed.

Context chips keep the same appearance in the draft and sent message. Click an image chip in a sent message to open its preview. In the integration list, connected services appear first; you can drag the scrollbar without closing the list.

Choose Integrations to open the compact service list. Keep typing a name to filter it. Connected services have a Connected tag; you can also mention services that have not been connected yet. Hover over a context item or service to see its description on the right. Use the button at the bottom of that panel to open the related section.

Click an item or use the arrow keys and Enter or Tab to insert a chip into your draft, like a selected element. Hover over the chip to reveal its removal cross; click its name to open the related section. Chips stay available when you restore a draft or edit a queued message. The back arrow returns to the context menu and Escape dismisses it. Selecting a mention does not send the message or connect a service. See Integrations and environment to connect one.

Open commands with /

Type / on Home, in Support, or in project chat to open Commands. Outside a project, only /sound is available; actions such as renaming a chat require an open project. Type a command name to filter the list, use the arrow keys to select, then press Enter or Tab, or click an item. Escape closes the menu. Choosing a command removes it from your draft without sending a message; the rest of your draft stays intact.

/sound switches between Turn sound off and Turn sound on. Sound is on by default. Your choice is saved in your account and applies to all projects when you sign in on another device. The same setting is available in the account menu beside language and theme. It controls chat sounds when the agent finishes, needs your answer, or encounters an error.

In project chat, other commands open the existing controls; complete the action there:

• Export code, Export database, Connect integration, and Configure device open the relevant project tab.
• Rename chat opens title editing for the current chat; Archive chat opens its confirmation.
• Manage access, Manage publishing, Configure GitHub, Make a copy, and Archive open the project's existing dialogs. Archive refers to the project.
• Contact support opens Support. Check project usage opens AI Gateway with this project already selected, including when it has no usage yet.

The archive-box icon marks Archive chat; the archived-folder icon marks Archive project. Available commands depend on your role and the selected chat.

Follow the work

When your request requires work beyond a direct answer, Amazi briefly describes what it is starting. Greetings, explanations, and requests to confirm understanding without acting receive a direct answer. Amazi skips a separate progress update when it would only repeat that answer.

Your messages have a neutral background. Amazi's replies appear as plain text without a message background. The sender's avatar, name, and time appear together above the message; reply actions and work details remain below it.

Your work summary remains visible while the completed reply is loading. Questions, selected answers, and requests for your input use brighter text than background agent activity.

Links in Amazi's replies stay clickable in the text and also appear as chips below the message. Repeated links appear once when both their name and address match. Links with different names remain separate even when they lead to the same address. A book marks a documentation page; an outward arrow marks another link. Amazi reads English or Russian product guidance in the language of your question, or in the language you explicitly request.

The work summary also records integrations you connected or skipped and environment-variable requests you saved or skipped. These actions remain in the conversation after the task finishes. You see integration names and variable names; secret values are never shown or copied with the summary. Documentation references in Amazi's replies open the corresponding guide in your interface language.

The copy button beside an agent reply copies only that reply. To copy progress updates, open the work summary and use its separate copy button.

Below a completed reply, credit usage and elapsed time have separate icons. When the reply includes saved file changes, the File changes header shows the commit's added lines in green and deleted lines in red. Expand it to inspect the files. These counts cover text changes; binary files do not contribute line counts.

Between updates, a wrench icon shows how many background actions Amazi has completed. The count updates while work continues and remains in the saved summary. Actions already shown as questions, progress updates, or image cards are not counted again.

Use the work summary to follow short updates about progress, findings, and verification. These updates describe the work in plain language. Image generation shows a loading indicator followed by the generated images, including design references. Previews remain available in the saved summary to people with access to the project. Amazi handles investigation, implementation, and verification in the same conversation. The final reply briefly explains what is ready and any next step that matters to you. Technical details are included when you ask for them. Verification screenshots stay internal unless you request them; files you asked Amazi to create can still be delivered as attachments.

After complex tasks, Amazi refreshes the chat's suggested next steps. Short questions and small changes leave the existing suggestions in place; you can explicitly ask to review or change them. Selecting a suggestion fills the message field so you can review or edit it before sending.

When you are at the bottom of the chat, new activity keeps the latest content in view. Scroll up to read earlier messages without being pulled down. Use the down arrow to return to the latest activity.

For request examples, see Writing effective prompts. Keep recurring preferences in Project instructions.

Answer questions

When Amazi asks for an answer without choices, use the composer inside the question card to type, attach files, or reference selected elements, then select Finish. An attachment can be sent without text. For several questions, use Next to continue. Questions with choices still let you select an option; when available, choose Other to write your own answer.

The answer field has the same controls as chat: attachment previews, image references, mentions, and voice input. Use its send arrow or Enter to advance to the next question or finish the answers. Text and files stay with their question when you go back and forward.

After selecting an option, you can add a comment in the field below it. Its background matches the reply preview. When you submit your answers, Amazi receives both the selected option and your comment.

Review questions and your answers in Amazi's work summary. Each question appears separately with its own answer, and selected choices stay visible. Typed answers, comments, and attachment-only answers appear consistently as separate messages below the choices, with the sender's name and avatar above the message. Attached files appear in a carousel inside each bubble. Open an image or text file to preview it, or download another file.

Share files and large references

Attach references, documents, images, or source archives and explain how they should be used. One submission supports up to 10 files with a combined limit of 50,000,000 bytes. Amazi can also deliver generated files as downloadable reply attachments without building an application.

See Files and attachments for larger downloads and delivery limits, and Generating images for visual assets.

Chats, context, and queue

Questions, integration and environment requests, and queued messages remain available in the same pending-actions block across Chat, Chats, and History, even when you open another conversation. Switching views preserves unsent answers, attachments, the selected action tab, and whether the block is collapsed. Answers go to the original request. A queued message shows a conversation link only when it belongs to another chat.

If a request from New chat enters the queue, it is marked New chat. Amazi creates its separate conversation when the request reaches the front of the queue. Each such request gets its own chat. Editing returns the request to the new-chat composer; deleting it before dispatch leaves no empty chat. Once created, the conversation is reused if delivery is retried.

Use the arrow at the upper-left corner of the selected tab's content to collapse questions and queued messages together. The compact button below the conversation opens the block again; answers already entered stay in place. A soft fade always separates the conversation from the message field.

Long questions, environment fields, and queues scroll inside the content area. The action tabs and collapse button remain in place while you scroll.

If editing a queued message fails, read the toast notification and try again. The queued request remains available and the current draft is preserved. The maintenance details dialog shows the remaining time alongside separate start and end cards.

Queue cards show one line of the message. Use the arrow on the right to expand the full text or collapse it again. References and attachments stay available; editing and sending always use the full message. Your message text keeps its normal size after sending, including paragraphs followed by Markdown separators.

Replies keep the referenced message in the queue. Editing restores the reply preview in the composer; sending normally or to a running task keeps the reply context and referenced attachments.

Attachments appear as file cards inside the queued message, with names, sizes, and image thumbnails. Open a card to preview an image or text file, or download another file, just as in a sent message.

Maintenance notices appear at the top of the workspace and beside the chat composer. Click the top notice to open the description and scheduled start and end times. Scheduled notices count down to the start; the active top notice counts down to the end. The end time updates if maintenance is extended. During maintenance, you can send a message in an existing project to save it to the queue, including for a new chat. Queued messages wait without being sent to the agent, and steering an active task is disabled. The queue resumes automatically when maintenance ends; you do not need to send the messages again.

While a task is running, any member with chat editing access can steer it with a queued message, even if another member started the task or wrote the message. The message keeps its author and attachments. In the work summary, its original send date and time appear beside the sender's name.

Each project can have several chats. Use a new chat for an unrelated subject and rename chats to keep them recognizable. Amazi manages long conversations automatically. Archiving makes a chat read-only and restoring reopens it. The chat input remains visible with an explanation inside. Composer action buttons stay in place when unavailable and become disabled.

If Amazi fails after accepting the first message in a new chat, that chat stays open with your message and the error. You can review the failure and continue in the same conversation.

When a new-chat request starts immediately, its chat opens and shows live progress without reloading the page. A chat created from the queue appears in the Chats list when its turn starts. Activity updates appear as they become available, including starting, stopping and completion, without a one-second batching delay. While Amazi is working, you can stop the task from the chat. The stopping indicator remains visible until the current work has stopped and its changes have been saved.

If Amazi reports that a request is too large for the model, split it into smaller tasks or provide large reference materials as files.

Completed replies show the recorded working time alongside credit usage. The time appears when the task finishes and stays visible as the chat updates, without reloading the page.

You can select Reply on an error message to add your own instruction, or Continue to resume. Both quote the visible error message; hidden technical details are not included in the quote.

A project runs one task at a time. Additional submissions enter the shared project queue with their destination chat and author. The full queue remains visible when you switch chats; dragging an item changes its position in that shared queue. Permitted users can review, edit, redirect, or remove queued requests before execution. Editing returns the request and attachments to the composer for revision and resubmission. Stopping a task prevents further work and saves the changes made so far in History with an Interrupted label. Wait for stopping to finish; a saved version may still contain unfinished work. Retry or Continue is sent as a reply to the failed assistant message, so Amazi receives the original failed request and attachments together with the recovery instruction. Completed work shows the task's total credit usage across its separately metered operations. View-only collaborators can read chats, while working with Amazi requires editing access.