Assistant

The assistant is the panel on the right side of the app, a conversational companion that knows your matter inside out and can help you draft, research, and edit it. It is specialised in IP law (patent drafting, prosecution rules, jurisdictional differences) and built specifically for the work that happens on this platform.

The panel is available on every screen: with a matter open it sits beside the editors, and without one it accompanies the global tools (Prior Art, Classifications), so you can ask questions while you browse. Collapse or reopen it with the icon at the right edge; the app remembers your choice.

You talk to it in plain natural language. It understands the same smart references you use elsewhere (claim 1, FIG. 2, element names) and replies with chips you can click to navigate.

What the assistant can read

The assistant has full read access to your matter. You can ask it questions about anything in it without copy-pasting context. It can pull up:

  • Claims (the full claim set, including dependencies and history)
  • Claim diagnostics (the same errors and warnings the claims page's issues panel counts, such as missing dependencies or no transition phrase. Like the panel, it checks your saved claims, not changes still awaiting your review)
  • Application sections (Background, Summary, Detailed Description, and so on)
  • Figures and reference-sign elements (the matter's figures, their pins, and the named elements they reference)
  • Bookmarked patents (any prior art you've saved to the matter)
  • Saved web links (pages you've added to the matter context)

See the References tab for a full explanation of how chips work for patents and web links.

Ask things like "do my dependent claims align with the description?", "summarise figure 3", or "what does US 10,123,456 say about heat dissipation?" and it has the data on hand.

Beyond your matter, the assistant can reach out to two external sources:

  • Patent databases (search the patent databases with the same backend the Prior Art page uses, then pull a patent's full record into the conversation: bibliographic details and abstract, specific claims or the whole claim set, targeted passages of the description, CPC/IPC classifications, and the cross-jurisdiction patent family. It can also explore the CPC classification scheme itself, searching it by topic to find relevant codes, or looking up any code to see what it covers and browse its place in the tree)
  • The open web (search the internet and fetch the actual page, not just a snippet, when it needs to read something carefully)

You can ask it to run a prior-art search, look up a standard, find a regulator's guidance, or read a competitor's product page.

For non-patent prior art, a live page only proves what it says today, not when it was first published. Ask the assistant to check the Wayback Machine: it can retrieve a page's full capture history — when it was first archived, when its content changed, and when anything on its domain first appeared — and it can read the archived content of a page as it stood on a given date, so you can see what a page said before your matter's priority date or compare what it said at two different dates. Capture dates are evidence for you to weigh, not a legal determination of publication date.

The assistant only opens pages that you or a search brought up: a link you pasted into the conversation, a search result, or a link on a page it already opened. It will not visit an address it made up on its own. This keeps text inside an uploaded document or a search result from talking the assistant into sending your draft claims to an outside site, which for unfiled material would be a disclosure you cannot take back. If you want it to read a specific page, paste the link.

What the assistant can change

The assistant can edit your matter, but it never does so silently. Every edit it proposes appears as a pending change that you review and accept (or reject), either:

  • via Accept (or Accept All) in the pending-changes box above the assistant input, or
  • inline on the affected claim, section, figure, or bookmark.

A dot on a page in the left navigation (Claims, Sections, Figures, or Sources) means that page holds suggestions you can review there. Claims and Sections count only the application you have open. The dot clears once you have accepted or rejected them all.

Changes the assistant can propose:

  • Add, edit, or delete claims
  • Edit application sections
  • Add or remove bookmarked patents, and save web links
  • Add, edit, or delete figure elements and pins
  • Generate a new figure or revise an existing one, from a description, an uploaded sketch or photo, or another figure (see figures)

Actions for the page you're on

Some pages show a short row of actions just above the assistant input, once a matter is open. Click one and the assistant starts on it straight away. It first gathers what it can from the page you're on, your matter and the conversation, then asks only for what it still needs: for a new figure, which view and what it should show; for an edit or a check, what you want done, offered as a short list of options ending with "or something else". It never guesses before a step that costs money (generating or redrawing an image, running a search) or changes your draft. A search always shows you its query and waits for your go-ahead; see Patents.

  • Figures: Generate figure and Check figures, plus Edit figure while a figure is open.
  • Claims: Draft claims and Check claims, plus Edit claim when a claim link has brought you to one claim (the address shows ?claim=). See Claims for what each one does.
  • Sections: Draft section, Edit section and Check description. See Sections for what each one does.
  • Patents: Search prior art.
  • A patent page: Compare to claims and Search similar.

A check lists the problems it finds first, then proposes the fixes as pending changes for you to review. An edit fixes wrong or missing reference numerals by moving the marks, which is free and leaves the drawing untouched; it only redraws the image when the picture itself is wrong.

The row stays available throughout a conversation. It steps aside while you type, while the assistant is replying, and while pending changes wait for your review.

Each action is also a command you can type: /generate-figure, /edit-figure or /check-figures followed by what you want, for example /generate-figure side view of the hinge, /check-figures claimed features or /check-claims antecedent basis. Give it enough detail and it starts without asking. Type / in the input to see every command. Asking in your own words ("make a figure of the hinge", "find prior art for claim 1") follows the same rules.

Attaching context with @-mentions

Type @ in the assistant input to open a picker. You can attach a specific claim, figure, element, section, source file, web link, or bookmarked patent directly to the message, and the assistant treats it as a first-class input rather than something it has to go look up. This is the fastest way to ask focused questions like "@claim 5, can we narrow this?" or "compare @figure 2 to @US10123456".

@-mentions are distinct from the inline chips that appear when you type "claim 5" in prose: mentions are explicit attachments, chips are recognised references.

You can also attach something without typing at all: right-click a patent, claim, figure, reference-sign element, section, source file, or web link anywhere in the app (or use its ⋯ menu) and choose Ask assistant. It opens the panel if it's closed and drops the same attachment chip into the input, cursor right after it, ready for your question.

Asking about selected text

Select any passage in a section, a claim, a patent's claims or description, or the Preview panel, and a small Ask assistant about this button appears above your selection. Click it to open the panel (if it's closed) and drop the passage into the input as a quote card, headed by a chip naming where it came from, cursor right after it so you can type your question straight away. You can trim the quoted text in the input before sending, same as any other text there.

The same button appears on the assistant's own replies, so you can quote something it said back to it, minus the origin chip since it's already talking about its own words. Clicking the origin chip on a sent quote opens the section, claim, or patent it came from; it only shows orphaned styling if that item itself no longer exists.

Memory

Memory is what the assistant carries from one conversation to the next, so you never have to re-explain how you work or what a matter is about. Open it with the memory button at the top of the assistant panel. A background process reviews each conversation shortly after you leave it and folds anything durable into memory, so if you ask the assistant to remember something, it acknowledges the request in the moment and the memory itself appears a little afterwards, not instantly.

Memory has two parts:

  • User preferences apply across every matter. Use them for how you work in general: the offices you practise before, your default filing strategy, and your house style for drafting claims. You can read, edit, or delete any entry, and add your own.
  • Matter memory applies only to the open matter, so it appears only when one is open. It is a single running summary of the matter: decisions made and why, findings you have checked (including ones you rejected), open questions, and any standing instructions for that matter, along with when it last changed. Like your preferences, you can edit or delete any entry and add your own.

Conversation history

Conversations are saved automatically. Use the history button at the top of the assistant panel to browse past conversations and jump back into one. From the history list you can also rename a conversation to something memorable or delete one you no longer need (right-click an entry, or use its more-actions menu). The + button beside the history button starts a fresh conversation at any time.

Every conversation belongs to either a matter or to you directly (a global conversation, started without a matter open). The panel shows the conversations for where you are: the open matter's list (with your global conversations beside it) in a matter, your global list outside one. Each matter remembers the conversation you last had open in it, and so does global mode: switching matters shows that matter's conversation, or a fresh one the first time, and switching back brings the previous one back. The history view shows everything in one list. Matter conversations carry a small badge naming their matter, and opening one reopens that matter with the conversation ready in the panel.

A global conversation becomes part of a matter in two ways: send a message in it while that matter is open (your first message claims it for the matter), or use Move into matter… on its menu, useful when exploratory research turns out to belong to a case. Either way the move is one-way: once a conversation is part of a matter it stays there, and it can never move between matters, so a discussion of one client's case can't end up in another's context.

Fullscreen and docking

The assistant is one panel that can sit docked beside your work or fill the screen. The expand button at the top of the docked panel takes the current conversation fullscreen; the Assistant entry in the sidebar does the same and keeps you in your open matter. To dock it again, open any page from the sidebar: the panel slides back beside that page with the same conversation open, so you can keep talking while you work.

Reviewed sources and patents

The globe and document buttons at the top of the panel list every web page and every patent the current conversation has reviewed, so you can revisit what the assistant read. A web source opens in a new tab; a patent opens in the patent viewer.

Model

Choose the model the assistant runs on, either from the picker at the bottom of the assistant panel or under Settings → AI. The picker lists each model by its own name. Claude Sonnet 5 is the default and handles most drafting work. Claude Opus 5.5 is the strongest: slower and more expensive per message, worth it for the hardest claim drafting and prior art analysis. OpenAI's GPT-6 Sol and the low-cost GPT-6 Luna are also available.

The choice is yours, not the matter's: it applies in every matter and outside one, and stays until you change it.

Reasoning effort

The control beside the model picker sets how long the assistant thinks before it answers: Low, Medium, High or Max. Click it to open the slider, then drag or use the arrow keys. The same slider sits under Settings → AI.

Higher effort is slower and costs more per message, because the thinking counts toward your daily AI usage. It pays off on hard work: drafting a claim set around tricky prior art, or checking support for an amendment. Medium is the default; Low answers quick questions and small edits faster.

Like the model, the setting is yours and applies everywhere.

Context and usage

The small ring under the assistant input, opposite the model picker, shows how full the current conversation is: how much of the model's context window (the most text it can take in at once) the last reply needed. It starts empty on a new conversation, fills as the conversation grows, turns amber at 80% and red when the conversation no longer fits. When it gets close, start a new conversation; the assistant's memory carries what matters across. Click the ring to see the exact count and, below it, your daily AI usage, the same meter as in Settings → AI.

Jurisdiction awareness

The assistant adapts to the active application's jurisdiction (US, EP, or PCT). It pulls in the matching drafting rules (35 USC §§101/102/103/112 for US, EPC Articles 52–84 and the problem-solution approach for EP) so the suggestions you get match the office you're filing in. Set the jurisdiction on the application and the assistant follows.