How to use AI Chat Console

Sending your first message takes seconds — no setup required.

Different AI models are good at different things. Click the model name near the message box to open the picker and switch anytime — even mid-conversation.

• Everyday questions & writing — fast, low-cost models handle most tasks well and use the fewest credits.
• Complex reasoning, coding, analysis — premium models think more deeply (they cost more credits).
• Vision — models with a 👁 icon can look at images you upload.
• Web search — models with a 🌐 icon can pull in live information.
• Privacy — models with a 🛡 shield support Zero Data Retention.

Sometimes a model is busy, down for maintenance, or retired by its provider. Normally you would see an error and pick another model yourself. Switch on automatic fallback and the app does that for you: your message goes to the closest match instead, and the reply says so.

It is off by default. Answering with a different model than the one you chose should be your decision.

Illustration: the badge still shows the model you picked; the dashed note underneath shows who actually answered. Your chat stays on your model — the next message tries it first again.

How the replacement is chosen

• It can do what your message needs. An attached image needs a model that can see; a message that uses tools, the assistant or skills needs one that can call tools; web search stays with the same provider so your search still runs.
• Same or similar price. Never more than 1.5× what your model costs, and cheaper is always fine. A free model only falls back to another free one.
• Similar size. A small, fast model is replaced by another small, fast one, not by the most expensive model on the list.
• Somewhere that is working. If a whole provider is down, a different provider is preferred. If one model was retired, a newer model from the same family is preferred.
• Your privacy settings hold. With Zero Data Retention on, only ZDR models are considered.
• A limited number of tries — two other models and about 40 seconds by default. After that you see the error rather than waiting longer.

What it never does

• Use your own API key. If your message was running on a key you added under Settings → API Keys, you get the error — the app never quietly moves your request somewhere else, and never onto your key.
• Switch after the reply has started. If a model fails halfway through writing, you keep what arrived.
• Retry a refusal. If a model declines a request on content grounds, the app does not go looking for one that says yes.
• Retry a message that is too long. A prompt over the model’s limit would fail the same way elsewhere — shorten it, or start a new chat.

Worked examples

The buttons beside the message box add extra abilities to your chat:

The Prompt Library

Not sure how to ask? Browse the Prompt Library — ready-made prompts with real examples of what each one produces. Open one, press Use in chat, and it lands in your message box with the right model already selected.

• Examples, not promises. A prompt can carry the picture, document, spreadsheet or audio clip it actually made, so you can see the result before you spend anything trying it.
• Save your own. Anything you keep retyping can become a prompt. Keep it private, share it with named colleagues, open it to everyone signed in here, or submit it to the public library.
• Fill-in-the-blanks. A prompt can declare {{placeholders}}; you are asked for each one when you use it, rather than having to edit the text by hand.
• Fork anything. Found one that is nearly right? Fork it, edit your copy, and the original is credited.
• Type # in the message box to pull up your prompts without leaving the chat.

Pointing at a file with @

Once you have uploaded something to a chat, type @ in the message box to list every file in that conversation — what is attached to the message you are writing, and everything uploaded on earlier turns. Pick one and its name drops into your message.

• It says which one you mean. With three screenshots attached, “@chart-2.png what is the trend here?” is unambiguous where “the second image” is a guess.
• It brings an older file back. Long conversations cannot carry every document in full, so older ones are summarised down to make room. Naming a file with @ pulls that one back in complete — including a picture from earlier in the chat, which the AI otherwise cannot look at a second time.
• Deleting the text undoes it. The mention is just the filename in your message; remove it and nothing extra is sent.
• It works right away. A file is nameable the moment you send the message you attached it to — you do not have to reload the conversation first.
• Files that have passed their retention window are listed but greyed out — the file itself is gone, so there is nothing left to bring back.

Pointing at a file in Draw, Video and Speech mode

@ works in every mode, not just chatting — but what it does depends on the mode, because each one can use a file for something different. You are only ever offered the kind of file that mode can actually use, so the list is never full of things that would do nothing.

• Reference images are nameable too. A picture you attach with the reference button in Draw or Video mode joins that conversation’s file list for the rest of your visit, so a second or third attempt can pull it back with @ instead of you finding it on your computer again.
• Speech mode: the mention marks the spot. “Here is this week’s update. @notes.md That is everything for now.” is spoken as your two sentences with the whole document read out in between. Put the mention on its own to have nothing but the document read.
• A document has to have readable text. If nothing could be pulled out of it — a scan with no text layer, for instance — Speech mode will say so rather than read out the filename.
• It is always optional. Every mode works exactly as before if you never type @.

Speech mode charges for the words it actually speaks, so a mentioned document counts towards the length limit and the cost of that clip — a long PDF is a long clip.

Bringing a large document or an image back into a reply uses more of the conversation’s context, and images cost more than text — so mention the file you actually need rather than all of them.

Ask for something to happen and the AI sets it up instead of telling you how. Say “remind me tomorrow at 10 about the ABC meeting” and a card appears in the reply with the time it worked out. Press Confirm and it is set. Ordinary questions are still answered as ordinary answers — nothing changes about normal chat.

Illustration: asking for a reminder in chat. The card shows what will happen and when; the green line is what it becomes after you press Confirm.

Try these first

What it can set up for you

• Reminders — one-off or repeating, delivered as a phone notification, an email, a chat message, or all three. Free; they never touch your credits.
• Calendar entries — “I have a meeting with ABC on Thursday at 11.” Anything already booked at that time is flagged on the card, and it can set a reminder a few minutes before.
• Background tasks — “every Monday check these pricing pages and email me a comparison.” The AI plans the steps, does the work and delivers the result, with the tab closed.
• Meeting briefs — “brief me before the ABC meeting.” Shortly before it starts you get a short prep note: what it is about, anything recent about the company or people, and what to be ready to say.
• Page watching — “tell me when their changelog changes.” Checked on a schedule; you only hear about it when something actually moved.
• Email — to yourself now or at a set time, or to one of your saved contacts with the unsubscribe link and compliance footer already handled.
• Contacts, workflows, modules and deep research — anything you have set up elsewhere can be started by asking for it.
• Questions about your own account — “what is on tomorrow?”, “how many credits do I have left?”, “is anything waiting for me?”

Illustration: the Assistant page. Reminders, meetings and scheduled tasks in one list, grouped by day, with anything overdue marked.

How confirming works

Quick ways in

Asking in a sentence always works, but two shortcuts skip the AI entirely — they are instant and cost nothing:

• /remind in the message box. Type the whole thing in one line — /remind call the bank tomorrow at 10 — and a small dialog opens with the time already worked out. Change anything you like, choose notification, email or chat, and set it.
• /agenda shows what is coming up in the next week, with a Done button beside each reminder.

Illustration: /remind reads the time out of the sentence — “tomorrow at 10” becomes a real date you can correct before saving.

• In the Chrome extension, type @remind to be reminded about the page you are reading — pick In 1 hour, This evening, Tomorrow morning or Next week, and the page title and link are kept with the reminder. @agenda lists what is coming up without leaving the panel.

Illustration: the browser side panel. @remind keeps the page you are reading for later, with its title and link attached.

Before a meeting

Ask to be briefed and the assistant schedules a short piece of research that runs shortly before the meeting starts — by default half an hour. It arrives as a notification and an email: what the meeting is about, anything recent it can find about the company or people involved with links, and a few things to be ready to say. It runs the AI, so it spends credits from your balance, and it says so on the card before you agree. Cancel the meeting and the briefing is cancelled with it.

Set your timezone first

“Tomorrow at 10” only means something if the app knows where you are. Open Assistant → Settings and check the timezone; everything the assistant schedules uses it. If your device disagrees with the saved setting, the page offers to switch.

Quiet hours, the daily brief and your calendar

• Quiet hours hold notifications overnight — they wait in your tray instead of buzzing. Emails you specifically asked for still go out on time.
• Daily brief is an optional email each morning: what is on today, anything overdue, and anything waiting for your confirmation.
• Calendar subscription gives you a private link to add to Google, Apple or Outlook Calendar, so assistant events and reminders appear beside everything else. Rotate the link any time if you share it by accident.

What it costs: reminders, calendar entries and organising things are free. Only work that runs the AI — a scheduled task, deep research, a module — uses credits, and the card tells you the budget before you agree to it.

Safety: only your own message can trigger an action. Text inside a web page, a PDF or a file you uploaded can never make the assistant schedule, send or spend anything by itself, and nothing that emails another person is ever sent without you seeing the draft first.

A chat answers you now. A task goes away and does the work — searching, reading pages, writing, building a document — and comes back with the result. It runs on the server, so you can close the tab, and it can run again every week without you. The Tasks page lists every template, the three planning modes and the five delivery channels.

Making one

Or skip the form entirely: say it in chat. “Every Monday check these pricing pages and email me a comparison” builds the same task and shows you a card to confirm — see Assistant & reminders.

Illustration: a task card and the timeline of one run — every step, how long it took and what it cost, with the delivered file at the end.

Worked examples

While it runs, and after

• Each run shows a timeline: every step, how long it took, what it cost, and the output it produced. A failed step can be retried on its own.
• Approvals are yours to set: approve the plan before it starts, approve the result before it is delivered, both, or neither.
• Each run has a credit budget. It stops when the budget is reached rather than spending on quietly.
• Anything a run produced — a PDF, a spreadsheet, an image — is attached to the run and to wherever it was delivered.

Emailing clients

A task can send its result to contacts you keep in Tasks → Contacts, imported from CSV or added one at a time. That side is deliberately careful: you add your business name and postal address first, every message carries them and an unsubscribe link, new senders have their mail reviewed before it goes, daily limits apply, and bounces and spam reports pause sending automatically. You can send through the site or connect your own mail server so it comes from your own domain.

What it costs: a run spends credits like a chat does — more for more steps. The budget you set is the ceiling, and Tasks will not start a run if your balance is below the site minimum.

A skill is a set of instructions for one kind of job — how you want invoices laid out, what your code reviews must check, the tone your newsletters use. Write it once and the AI follows it whenever a matching request comes up, without you pasting the same briefing again. 20 are built in and install in a click; the Skills page lists them all with the request each one answers.

How the AI picks one

• Only a one-line description of each skill is sent with your message. The full instructions are loaded only when one is actually used, so having a shelf full of them costs you almost nothing per message.
• Type /skill-name to force a particular one.
• Turn on Always on a skill's card and it applies to every message.
• Otherwise the AI chooses by itself from those descriptions — which is why the description should say when to use it, not just what it is.

Getting them

Illustration: the tester. One message, two answers — with your skill and without it — plus the tokens, credits and time each took.

Two ways to start

• Ready-made — Skills → Library has 20 working skills across Writing, Marketing, Business, Development, Analysis and Productivity: email replies, meeting notes, invoices, code review, SQL, proofreading, plain English, bug reports, test cases, spreadsheet formulas, interview guides, decision briefs, study notes and more. Search it, install one, and it works in every chat straight away. You can edit it afterwards like your own.
• From a shape — + New skill now asks what kind of skill you want first: a house format (invoices, reports), a review checklist, a reply in your voice, a detail extractor for messy text, a document with fixed sections, a rewrite to your tone, or answers from your own rules. Pick the nearest one and the name, the description and a first draft of the instructions are already written — you edit rather than start at a blank page. An empty skill is still one click away.

A skill that works, in three lines

The description is the part that matters most: it is the only text the AI sees when deciding whether a skill applies. “Use when…” belongs in it.

Keeping them current

A skill imported from GitHub can check for updates. You see the changes line by line before applying them, and the version you were using is kept so you can roll back. Skills can also be shared to the site library for other people to use, once an administrator approves.

Skills change how the AI writes; they never grant it new powers. A skill cannot give itself tools, reach your files, or override the rules the site runs under — and a task can be told to follow specific skills too.

Switch to Draw mode with the image/draw button beside the message box, then describe what you want to create.

Switch to Video mode with the video button beside the message box (or type /video), then describe the video you want.

Switch to Audio mode with the audio button beside the message box (or type /audio), then type the text you want turned into speech.

Modules are guided, fill-in-the-blank AI templates that turn a task like “write a Facebook post” into a short list of questions instead of a blank prompt box. Answer them and the AI does the rest. 265 are included, across 27 categories, and you can build your own.

• Each module can restrict which AI model it runs on — if it needs one your plan doesn't include, you'll see that clearly before submitting rather than after.
• Image modules use the same Draw pipeline and credit cost as generating an image manually.
• A module that's still mid-way through its questions can be cancelled at any point without sending anything.

What is in the library

Each category is searchable, and the library page has a filter chip per category:

• Daily & weekly routines (10) · Planning & focus (9) · Meetings & prep (9) · Inbox & messages (8) · Watch & alerts (8) · Team & management (8) · Freelance & clients (8)
• Writing & editing (12) · Marketing & ads (20) · Social & short-form (14) · SEO & content (12) · Sales & CRM (8) · Customer support (8) · Business & operations (16)
• E-commerce (11) · Finance & analysis (9) · Legal & compliance (9) · Career & HR (10) · Teaching & study (11) · Learning routines (6) · Research & data (9)
• Developer & technical (10) · Creative & media (9) · Property (7) · Bangladesh & regional SME (8) · Home & family (8) · Personal & everyday (8)

Modules for work that repeats

Most of the library is for one-off jobs. The routine categories — daily and weekly routines, planning, meetings, inbox, watch and alerts, team, freelance, home and learning — are written for repetition instead: each one reports what changed since the last run rather than restating everything, and says so plainly when nothing did. A morning brief, a standup note, a competitor watch, an inbox triage, a weekly review.

• Any module can be given a schedule — hourly to monthly, in your timezone — and then runs in the background and delivers the result without you opening the app.
• Every firing re-checks your plan and your daily run limit, so a lapsed entitlement stops the schedule rather than quietly spending.
• A handful are multi-step packs: several runs chained together, each reading the one before it — research, then draft, then a finished document. 17 of them, and your plan sets how many steps you may chain.

A project is a workspace for one job — a client, a course, a product launch. Chats inside it share the same instructions, notes, memory and files, so every new conversation starts already knowing the background instead of you re-explaining it each time.

• Project memory is separate from your account memory. Facts learned inside a project stay in it, so nothing from a client project leaks into an unrelated chat.
• Each project chooses how the two combine: merge both, use only its own memory, or use only your account memory.
• Pinned notes hold the current state of the work — decisions made, things still open — and are sent with every message in the project.
• Its own model. A project can pick the model its chats start on, so research work opens on a strong reasoning model while routine drafting opens on a cheaper one. Tick Always use this model in this project and opening any chat there switches the composer to it, not just new ones — you can still change it for a single message.
• Knowledge files are searched by relevance, so only the passages matching your question are included rather than whole documents.
• The sidebar stays where you put it. Picking a project filters your chat list to that project; opening a project chat from All chats does not — the chip above the composer tells you which workspace the current chat belongs to, and Close leaves the project without touching the list.
• Sharing and collaboration. Send a read-only link to a single conversation, or invite someone in as a collaborator to work alongside you in a chat or a whole project. Anything a collaborator sends is billed to the account that owns the workspace, so invite people you trust and remove access when the work is done.

Memory carries useful facts about you from one chat to the next — your role, your preferences, how you like answers written — so you stop repeating yourself. You decide what it keeps, and you can switch it off for any single conversation.

• Two kinds of memory. Facts you add yourself, and facts learned automatically from your chats. Both are listed in Account → Memory, where you can edit or delete any of them.
• A switch on every chat. The memory button beside the message box turns memory off for that conversation only — nothing is recalled into it and nothing is learned from it. Every other chat carries on as normal.
• What new chats start as is your choice, in Account → Memory. Leave it on, or have every new chat start memory-free and switch it on only when you want it.
• Auto-memory can be turned off separately if you would rather add facts by hand. Recall still works; only the automatic learning stops.
• Recall depth is adjustable — send everything remembered, or only what is relevant to the current message. The trade-off is thoroughness against token cost.
• Turning memory off entirely, in Account → Preferences, hides the per-chat switch too, since there is then nothing for it to control.

The extension puts the same chat in a side panel next to whatever you are reading, so you can ask about a page without copying it into another tab.

• Model picking, web search, reasoning and document generation all work the same as on the web app.
• Images and video can be generated from the panel too.
• Projects and the per-chat memory switch are available there as well, so a chat started in the panel behaves exactly like one started on the site.

Shortcuts are your own saved prompts — instead of retyping the same request every time (a writing style, a report format, a recurring image size), save it once and insert it in a couple of keystrokes.

• Shortcuts aren't just for text — they can also carry a saved image/video/audio configuration, so / can jump straight into Draw, Video or Audio mode with your usual settings pre-filled.
• Admins can publish preset shortcuts for everyone — those appear in the same / picker alongside your own, marked as presets.

The service runs on credits. Each message uses a small number of credits depending on the model — cheaper models go much further.

• Plans run on a 30-day cycle counted from the day you subscribe, not from the 1st of the month. Each cycle grants your credit allowance.
• An invoice is issued at the start of every cycle, before you pay it — find it under Account → Billing, with a PDF to download.
• Nothing is auto-charged to a stored card. You renew each cycle yourself, through whichever payment method you prefer.
• If a cycle goes unpaid there is a short grace window with full access, after which the account moves to the Free plan. Nothing is deleted — chats, projects and any credits already in your balance stay.
• Credit packs are one-time top-ups that never expire.
• Your current balance is shown in the app; top up anytime from Account → Buy Credits.
• Daily fair-use limits keep the service fast for everyone — see the Fair Usage Policy.
• For renewals, invoices, changing plan and what happens when a cycle goes unpaid, see the Billing guide.

Your prompts are sent to the AI provider you pick, used to generate a reply, and returned to you. Saved chats stay in your account until you delete them.

• Delete any chat — or your whole history — at any time.
• Export your data in several formats before deleting.
• API keys you add are encrypted and never displayed in full.