Start with a question. Build from what comes back.

Signed-in users can choose Get my brief and select one read-only lens. Robbie always starts with the small saved-work snapshot; a calendar, priority-mail, or What changed lens is used only when you select it for that one brief. It does not send, schedule, create tasks, or change connected apps.

What changed makes one bounded, cited search and returns at most three material updates. After that brief, choose short topics to focus future updates, such as a project area or market. The topic list is encrypted in your account and applies only to a later What changed brief. It does not start a feed, background watch, or additional search.

Robbie can be installed from your browser’s menu as an app. On iPhone or iPad, use Share then Add to Home Screen; on Android or desktop Chrome, use Install app or the install icon in the address bar when it appears. The installed experience opens as its own window and keeps the same signed-in workspace. Conversations, connected sources, and model requests remain online and are never cached for offline access.

Where your installed browser shows Share with Robbie in its system share sheet, you can hand Rob a page title, selected text, link, or supported file from another app. Robbie opens it as an editable private draft in Rob's composer. The shared material is not sent to a model, search, or connected app until you review it and choose Send.

Ask for a PowerPoint presentation or slide deck and Robbie creates an editable .pptx file you can download from chat or Build & run. A deck can include an opening subtitle plus opening, section-break, content, and decision slide structure, concise slide text, speaker notes, metric cards, compact charts, comparisons, timelines, workflows, decision matrices, attributed quotes, and owner-selected images. Give Robbie the audience, the decision the deck should support, an approximate slide count, and the evidence that belongs in the story.

Open Present to rehearse the delivered deck. Deck map gives you a visual overview and can find a slide by title or content. Check deck points out missing framing, notes, evidence, density, and a closing move. Presenter view keeps the current slide, talk track, next slide, pacing, and elapsed time together; its timer can run independently of Auto play. Audience display opens a synchronized, local second-screen view so the audience sees only the slide while the presenter keeps notes and controls; pause the audience screen during questions or a transition, then resume at the same slide. Download Notes for a speaker script, Handout for a clean local HTML reading copy that retains metrics, chart values, comparisons, and timelines, Slides PDF for a polished audience version, or Presenter brief for a printable per-slide talk track. All four exports are built locally from the deck already in your browser. Edit deck changes structured copy, notes, and visual data, then rebuilds a separate editable PowerPoint while the original download stays available. In the deck editor, select a generated image or add a PNG, JPEG, or WebP logo, chart, or photo from your device. You can also describe a visual directly in the deck editor to generate one presentation-sized private image. It remains in your private gallery and is embedded in the rebuilt PowerPoint.

Use the Goal picker above a saved conversation to link it to one of your saved goals. Robbie then receives a bounded private packet of that goal, its linked tasks and notes, and active linked decisions on later turns. This is an explicit chat-level choice, separate from automatic memory recall; change it or clear it at any time. The link never changes the goal or its work, and private goal context is withheld while text sanitization is on.

In a signed-in standard chat with personal memory on, type a request, open Add then More options, and choose Preview saved memory. Robbie checks the current draft only after that click and shows the saved titles and matching reasons that may be useful. Edit the draft and the receipt clears, so it never stands in for a new request.

The preview does not display note bodies or show attachments, connected apps, Team work, tools, or a future answer. It is unavailable for guest, sanitized, and saved-agent chats. The reply itself still shows its bounded context receipt after it completes.

Open Apps to see which connections are available. When Google Drive is enabled, connect your account, choose Google Docs, then select up to four for your current chat. Robbie reads their text and can cite the original documents. Select only what the conversation needs, and remove the selection when you are done.

Google Workspace can separately connect Gmail, Drive and Calendar with only the scopes you select. In Apps, Gmail and Outlook each let you choose up to five messages from one saved search and ask Robbie to prepare an encrypted action review. It labels proposed work as an action, informational, or uncertain; no task is saved until you select the exact proposals to add. Its read-only Drive connection can bring Google Docs, PDFs, Word documents, spreadsheets, presentations, and supported text files into a conversation as bounded reference text. Calendar reads stay read-only. If your operator enables the reviewed Calendar creation feature and you reconnect with that capability, Robbie can prepare an event in chat or Apps. The event is stored as an encrypted, one-use review first; inspect every field and type CREATE EVENT before Google receives it. If Robbie cannot confirm the provider result, it marks the event uncertain and will not retry it.

Microsoft works similarly: choose individual OneDrive or SharePoint files before Robbie can read them. Selected PDFs, Word documents, spreadsheets, presentations, and supported text files are converted into bounded reference text; Robbie does not receive general access to your drive and cannot alter the selected files.

Google sign-in does not grant document access. This connection currently supports selected Google Docs in chat, including document tabs; embedded images and other Office formats are not part of that read. Connected documents cannot yet be used with sanitization, Canvas execution, or saved reasoning runs. Compatible remote MCP services remain available separately for approved tool actions.

When Private text channels are enabled, open Connections and link an SMS or WhatsApp number with a short-lived code, then text Robbie directly. Each linked number has its own encrypted Robbie conversation and uses the same bounded personal context as Rob. Robbie verifies the messaging provider before it reads a text.

Text is meant for quick thinking, drafts, and context-aware help. Use FIND: launch brief to search your encrypted saved work without calling a model, or RECALL: launch partner to deliberately recover matching snippets from your own earlier Robbie conversations. BRIEF gives a read-only snapshot of saved tasks and explicitly due relationship check-ins; REMEMBER: reference saves one private reference note; TASK: title or TODO: title captures one task, while DECIDE: scope | choice | why captures an explicit Decision Ledger record without a model. Use PREP: Avery Chen when you want Luna Pro to prepare a thoughtful brief from one saved People record. Robbie never infers a task or decision from a normal message, local capture does not invoke a model, and replayed provider deliveries cannot create duplicates. Use RUN: Daily focusto start a routine already configured for Robbie or a saved agent; its exact name is required and a delivery retry recovers the same private Run. Use WORK: compare my launch options to queue one private background task, then DIRECTION: lead with the decisionto refine the newest active task started from that same linked number. Robbie uses a direction only at a future safe checkpoint; it never changes an in-flight model call, approves an app read, or takes an external action. Use STATUS for a short update, TRACE for its compact owner-visible model, usage, cost, and tool-activity receipt, RESULT for its newest completed answer, or STOP to cancel the newest active text task. When an operator explicitly enables an outbound sender for that channel, a settled WORK result is sent back once to the same verified number; otherwise it remains in the encrypted Robbie thread and Runs. Ordinary text cannot browse, access connected apps, upload files, send messages, make calendar changes, or run code. If you have separately linked a Mac with remote access and its Robbie app is open with Local Control ready, MAC: summarize files in Downloads can queue an explicit task on that Mac. The Mac may use the local permissions you granted it. Text MAC STATUS,MAC RESULT, or MAC STOP to check it or cancel before pickup. When outbound text delivery is enabled for that line at launch, Robbie sends the finished result back once. Continue in the web workspace for task details and any connected-app approvals.

Supported providers can reuse previously processed prompt context. The small cache indicator reports cached input tokens when the provider supplies that information. Cached tokens still belong to the model’s context; they are reused processing, not extra context capacity or a reused answer. Cache pricing and retention vary by route, and an absent readout means usage was not reported.

In a signed-in chat, expand a reply's Provider trace to see the model and upstream provider reported for each response step, timing, reported token and cache use, cost, and tool-call timing. Open a completed tool card for its bounded, redacted request and returned data, or copy the combined execution receipt for your own records.

A model's visible reasoning appears only if that provider actually supplies it. Robbie does not expose hidden chain-of-thought, prompt and memory packets, attachment bytes, credentials, headers, or raw provider envelopes. On a completed signed-in reply, you can choose Check answer for a bounded independent advisory check. It compares only your visible request text and that reply for coverage and concrete support; it does not edit the answer, inspect hidden reasoning, or receive attachments, memory, tools, or chat history.

Use the model picker to choose a route, or configure each agent’s model when building a team. Robbie checks supported file inputs and applies the route’s usage limits. Usage throttling may substitute the configured fallback after an allowance is reached; see the terms for that policy.

Turn on Think when a decision, plan, or analysis needs deeper work. For compatible text, image, and document inputs, Robbie routes that turn to GPT 5.6 Luna Pro, which has the same displayed per-token rate as Luna. Audio and video stay on the selected full-multimodal route because Luna Pro does not accept those inputs.

Goals hold the longer horizon: an outcome, optional target date, topics, and an active, paused, or completed state. Tasks remain the practical next steps. You can ask Robbie to read or update a goal, but a goal never creates tasks, starts an agent, schedules a reminder, or changes a connected app on its own.

Use a goal when you want Robbie to understand what the work is for across multiple conversations. Add a few topics to make the goal easier to find later, then ask Robbie to break it into a plan or review progress when you are ready. The Overview’s Follow-through lens can surface an active goal with no open attached task or an active decision with no linked goal. It reports only those saved facts; a missing link can be intentional. Review with Robbie hands over exact IDs for discussion and prepares changes only when you ask. Open a project and choose Schedule review when you want a recurring check-in. It stays bound to that exact project. You can choose an in-app handoff, a private Robbie result, or a saved-agent run; automatic work uses a fixed project snapshot and never gains connected-app, attachment, or external-write access.

Use Decisions when you need to preserve a choice across changing conditions. Capture the choice, why it made sense, alternatives, and the assumptions it depends on. You may set a review date; it becomes a private cue in Personal Focus and Personal Continuity, where you can reopen the record and decide whether to keep, change, or supersede it. A date never starts a model, checks an app, or changes the decision on its own.

Personal routines are owner-authored prompts with a local schedule: for example, a Friday reset, a quiet end-of-day review, or a monthly ideas check. Their names and prompts are encrypted at rest. At the selected local time you can choose a generic in-app handoff, have Robbie prepare a private result, or run one of your saved agents. Robbie’s routine agent is an ordinary saved agent that you can inspect and refine. Automatic runs use a fixed relevant-memory snapshot and never use connected apps, attachments, external writes, or messages. Choose General, Research, Think, or Create for a handoff. A routine configured for Robbie or a saved agent can also be started immediately with Run now; you can edit, disable, or remove it at any time.

Open Memory to save useful project context, preferences, people, and decisions. Use Personal profile to describe how Rob should work with you over time, and add short personal anchors for facts or preferences that should remain easy to review. All saved context is encrypted at rest and is treated as reference data, never a command or permission.

When a conversation establishes a useful, non-sensitive preference or project fact, Robbie may offer an editable Memory suggestion. It is only a proposal: review it, change it, save it, or dismiss it. Memory maintenance highlights expired and older entries so you can renew, revise, retire, or confirm them without changing their text. Robbie normally selects a small relevant set for a new request and shows a receipt beneath its reply. Memory preferences contain the uncommon controls for recall, conversation continuity, and semantic matching. When semantic matching is available, its encrypted index can match meaning when words differ; contact receipts are excluded, and you can erase the index from the same page. Explicit chat references, linked projects, and selected working agreements remain separate, visible choices.

For work that needs context even when the next prompt uses different wording, use References beside a saved chat’s Goal picker. Select up to six active notes as that conversation’s explicit working set. Saved background agents also offer Working notes: choose up to three active notes for one agent. On a later run, those deliberate references replace fuzzy note recall for that agent; the resulting encrypted snapshot is fixed at launch, and the run receipt shows titles only. Change or clear a selection whenever you want; it does not edit the notes or change a goal.

Saved agents can also carry up to three selected reusable skills, frozen at launch, and an optional Persistent playbook: a short, owner-reviewed operating record for future runs. Use it for durable preferences such as evidence standards, format, and what uncertainty to flag. Robbie snapshots it with each run; it does not rewrite itself, create a task, or grant permission to send, schedule, or change anything outside the run.

Sanitization is off by default. Enable it when you want supported personal details in text masked before the model request. It changes what the model sees, so review names, relationships, and important conclusions. Sanitized turns pause profile, anchors, saved-note recall, active decisions, and explicit chat references; they also use the isolated sanitized route when it is configured. It is not end-to-end encryption or a guarantee that every sensitive detail is detected.

On browsers that expose native speech recognition, the microphone in the composer turns speech into editable draft text. Robbie does not receive or store an audio recording from dictation; the browser’s speech-recognition service has its own processing policy. Review the transcript before sending it. A reply’s speaker button uses the device’s local speech synthesizer after you press it; it does not send the reply to a separate Robbie text-to-speech service.