How Does the AI Agent in Rideful Work?

Rideful has an AI agent built into the app — a composer on your dashboard and a chat panel in the editor. Describe what you want and it plans, builds and checks the deck, then edits the slide you point at rather than rebuilding everything.

It is one agent you meet in two places. On the dashboard it settles what a new deck is born with, or finds the deck you mean, and hands the work over; inside a deck it does the work and keeps working with you. This page is what you see while it runs, and what you can do to it.

Two Places You'll Meet It

On the dashboard

A composer sits above your deck grid at rideful.io/slides, on Home and inside any folder. Describe the deck you want and it settles only what a deck cannot be created without: it derives the name from your request, puts the deck in the folder you're in, and uses 16:9 unless you say otherwise. The one thing it asks about — and only if you have any — is which of your own brand templates to build on. It designs nothing and holds no slide tools.

It writes what you agreed to brief.md, creates the deck, and opens it with the build already running. If you leave a conversation half-answered, it's still there when you come back the same day.

Ask it for something on a deck you already have — “animate my PowerPoint”, “add a summary slide to the Q3 review” — and it finds the deck, asking which one only when several fit, then opens it with your request already handed to the agent there. The dashboard conversation carries on as the deck's chat; the deck's previous chat is kept but no longer shown, as with Clear chat. If an agent is already working on that deck, or waiting for your approval there, it tells you and links the deck instead, so two runs never edit the same slides.

In the deck

The chat panel on the right of the editor picks up from that brief and starts building. It's also where every later change goes: describe what you want different and it changes that.

The dashboard composer is desktop only. The deck page has no chat panel on mobile, so a handoff would land somewhere that can't show the run.

The Chat Panel

Open by default, on the right of the editor. Drag its left edge to resize it — anywhere between 320 and 600 pixels, and it remembers the width. The × in the header closes it; a chat button brings it back.

The composer reads Ask anything…. Enter sends, Shift+Enter adds a line, Esc clears what you typed. Below it the panel shows what the agent can see: the slide you're on, and how many elements you have selected. That context goes with your message, so “make this bigger” means the thing you selected.

The Plan

Anything bigger than a one-line change starts with a plan. It appears above the composer as Agent Plan with a completed-of-total count, each step pending, in progress, or completed and struck through. It's a read-out, not a form — there is nothing on it to approve, and it hides itself a couple of seconds after the last step finishes.

The agent does wait for a go-ahead before it builds a whole deck. Say yes, or say what you want different first.

The Subagents

The agent delegates the work. A designer, a reviewer or the researcher it launches is one line in the chat — what it is for, how long it has run, its latest step, and the brief it was given — and the work itself is a conversation of its own, which the line opens.

• Designer — puts elements on a slide: text, shapes, lines, charts and, when its brief asks for one, a picture it finds or generates, styled to the palette and fonts the plan settled. This is the one whose changes stop for your approval.
• Reviewer — read-only. It screenshots a slide and reports what's wrong with it, and on a deck of more than one slide it makes a final pass across all of them together, looking for titles that drifted, colors that stopped matching, margins that moved.
• Researcher — searches the web when the deck needs facts the agent doesn't have, and writes a report with numbered citations and a sources list. It runs on its own, one at a time: meanwhile the agent sets the design system and theme and can talk to you, and it builds the slides once the report lands. The report is a deck file under research/, named for its topic and day, so later conversations read it rather than search again. Research report under its line opens it.

Approvals

Auto-approve tools, in the composer's + menu, is on by default: the agent and its designers change your deck without stopping to ask. Turn it off to review each change first. Then, before the agent or a designer changes anything in the deck — the theme, a slide, an element — the run stops and asks. The card says what it wants to do, and Show details reveals the exact call. Three buttons: Approve, Reject, and Auto-approve, which approves this and everything after it, releases every other card already waiting, and lets the rest through from their next step on. Enter approves, Esc rejects, Shift+Enter auto-approves. While a card is waiting the composer is disabled and reads Approve or reject pending tools… — that's the run waiting for you, not a hang. While the agent's own card waits, designer reports and the messages you send wait behind it.

The switch belongs to this deck and to you rather than to the browser: every tab and device you open the deck in shows the same switch, and your other decks are unaffected. Turning it back on also answers the cards already waiting here — including any that parked while the tab was closed, which resume on their own with nobody watching. Clicking Auto-approve on a card turns it on too. Turning it off gates whatever asks next; a step already running ungated finishes that way.

Files and Reasoning

The files that describe the deck sit in a narrow column of file icons at the right edge of the editor: rules.md, what the agent must follow; brief.md, written on the dashboard; design-system.md, the palette, fonts and margins the agent writes before it builds; and, below a divider, each research report, with a magnifier icon. They belong to the deck, not the conversation, so clearing the chat keeps them; click one to read or edit it (see Deck Files).

The plan is not a file; it lives in the Agent Plan strip.

Where the model reasons before it answers, that reasoning is folded into a collapsed Reasoning block. Open it to see how it got there.

What It Remembers

The agent remembers you across conversations and decks, so a preference you state once carries over to the next deck. It keeps two things. A short profile: what you do (teacher, student, consultant, sales and so on), your education level if you teach or study, your field (a subject or an industry) and the language you work in. And notes, one fact each: something about you, a preference (“never use emoji”, “our house style is Inter, navy and teal”), work that spans decks (a course, a weekly team update), or where your materials are (the template you always start from). Long material you hand it, like a style guide, is kept with its note and opened only when the work needs it.

Every save is a line in the chat: Remembered: No emoji on slides,Updated: … when it replaces a note,Forgot: … when it deletes one,Noted: teacher · secondary · biology for the profile. Recalled: … means it opened a note's long form before working. The agent on the dashboard and the agent in the deck both save; the designers and reviewers read your notes too, so a preference holds on every slide, but only the agent changes them. The researcher never sees them, since its searches go out to the web. What you ask in the conversation always wins over a note.

It saves only what you said yourself. Each save has to quote one of your own messages, so text from a web page, a research report, an imported deck or a deck file cannot become a note. It does not save anything that is only about this deck (that goes inrules.md ordesign-system.md), the deck's topic or content, anything about other people (your students, your colleagues), client names or figures unless you ask it to remember one, grades, health, family or contact details. A note holding an email address or a phone number is refused.

To see or change what it knows, open Settingsfrom the sidebar and go toWhat the assistant remembers. Edit your profile, delete any note, or Forget everything, which deletes every note and clears the profile. You can also ask in the chat: “forget that” deletes the note, and “what do you remember about me?” gets an answer from what is saved. Once you delete something, an open conversation cannot save it again from what you said earlier, and the agent's next turn no longer sees it. Notes are yours: on a shared deck each person's conversation uses their own, and Claude or ChatGPT connected over MCP keep their own memory.

Stopping a Run, and Runs That Outlive the Tab

Stop generating — the square button where Send was — ends the run. It changes to Stopping… and finishes the step already in flight before it halts, so nothing is left half-written.

You don't have to sit and watch. The run continues on the server after you close the tab, and the panel rejoins it when you reopen the deck. Two things follow from that:

• A run is already in progress… in the composer means another tab or device is holding this deck's run. Finish or stop it there. A run whose browser died stops blocking on its own after about 16 minutes.
• If a banner says the run hit its time limit, everything it finished is saved and Continue where it left off picks it back up. Run length is capped while Rideful is in early access.

History, Clearing the Chat, and the Daily Limit

Chat history is per deck. Opening another deck opens its own conversation, and clearing one leaves the others alone. Clear chat in the + menu starts a fresh conversation — your slides and deck files are untouched, and the old messages are kept, they just stop being shown.

There is a daily allowance. Check usage in the same menu opens a bar above the composer showing how much of today's you have spent before you reach it; the bar also opens by itself once you have used 80%. If you do reach it the composer is disabled and a banner gives the reset time, midnight in your time zone; at the reset the banner clears and the composer opens again, without a reload.

Undoing What the Agent Did

Agent edits go on the same undo stack as your own, so Cmd/Ctrl+Z undoes them exactly the way it undoes something you did by hand, and redo works the same. Nothing the agent does is outside your reach.

What It Can't Do

• Search the open web for images. It finds photos only on Pixabay and Wikimedia Commons, which state a licence for every file, so it can credit what it places. For any other picture, paste it into the chat or give it a link, and it places that. See Add Images.
• Design from the dashboard. The composer above your decks only creates a deck or hands a request to one. How many slides, the narrative, the visual style — all of that is settled in the deck, by the agent in the chat panel.

Related

• Create with AI — What to say to get the deck you want
• Troubleshooting — A stuck approval, the daily limit, a run that seems to have stopped
• Add Images — The one thing the agent can't add for you
• Agent Workflow — The workflow an agent follows, over MCP or in the app