Start a deep research run
Put Deep research: in front of your question in Auto and send it.
✦ Tutorial · Beginner
Run deep research and check its sources
Start a deep research run from chat, ask a question it can answer, follow the run, then check each numbered source before you use the report.
Start tutorialYOUR OUTCOME
By the end you will have run a deep research question from chat and checked the sources behind the claims you plan to use.
Start a run from Auto on a computer or the phone Write a question the searches can work with Check each numbered source before you rely on it
5 min
Ask a question research can answer
Name the place and the comparison you care about, so the searches it plans stay on your problem.
Follow the run
Read the plan card as it moves through planning, searching, reading and writing.
Read the report and check its sources
Match each numbered citation to the Sources list and open the pages behind the claims you plan to use.
When a run fails or comes back thin
Read the failure message, then try again later or reword the question.
What each lesson covers
The tutorial, lesson by lesson.
Start a deep research run
Put Deep research: in front of your question in Auto and send it.
Deep research runs from the normal chat in Auto, on a computer or on the phone. Type /deep in the composer and pick /deep-research, click the Research chip when an empty chat shows it, or write a request such as deep research on heat pump running costs. The command and the chip both put Deep research: in the composer, and you type the question after it. Oppermind then searches the web, reads the pages it finds and writes a report with numbered citations, all in the same conversation. It works from the question text, so files attached to that message aren't read. In Workspace Agent 2.0 and Code agent the same text asks the agent to run deep research and pass the report on word for word, but neither step is forced, so stay in Auto when you want a run and the report exactly as written.
Practice. In an empty chat in Auto, pick /deep-research, or click Research if the chip is showing, finish the line with a question from your own work, and send it.
Deep research: How do heat pump hot water systems perform in cold winters like Launceston's, and what drives their running costs?
Ask a question research can answer
Name the place and the comparison you care about, so the searches it plans stay on your problem.
Deep research starts by turning your question into several web searches, so what it finds depends on the words you give it. Name the place and the conditions, like Launceston winters or a small health centre. Say what you are comparing, and give a time frame if recent information matters. One clear question works better than a list of five, and a second run can take the next one. If you want academic sources, say so with words like peer-reviewed or literature review, and it leans towards journals and research databases. It reads public web pages, so it won't see your own files, pages behind a login or your organisation's internal figures. Bring those into the same chat afterwards with an ordinary message.
Practice. Take a vague question from your own work and rewrite it so it names a place, a comparison and a time frame. Run both versions and compare the Sources lists.
Deep research: For a small community health centre in northern Tasmania, how do heat pump and gas hot water compare on running costs and winter performance?
Follow the run
Read the plan card as it moves through planning, searching, reading and writing.
After you send, a plan card appears under your message and ticks off each stage. Starting deep research and Planning the research come first. Searching the web shows how many searches it is running, and Following up on gaps appears when it goes back for more. Reading counts the pages it is opening, and Writing the report from counts the ones it could read, so the second number can be smaller. On a computer the finished card reads Completed with the number of steps, and on the phone it reads Task Complete. A small deep_research card under it ends with OK or FAIL. The report arrives as the reply in the same conversation and stays after you reload the page. When a run fails, the reply says what went wrong. On the phone the card usually reads Task failed, with Failed under the step that stopped it, but a card can still look finished after a failure, so go by the reply. Lesson 5 covers what to do next.
Practice. Run a question of your own and expand the plan card. Note how many searches ran and how many sources the report was written from.
Daniel sends his question and watches the card under it.
Read the report and check its sources
Match each numbered citation to the Sources list and open the pages behind the claims you plan to use.
The report puts numbers like [1] and [2] after the claims they come from. At the end, Sources lists every page it read under the same numbers, each linked to the page. Oppermind removes any citation that points at a page it didn't read and says so, but a number only tells you which page the report credits for the sentence, not that the page is right or that it says the same thing. Open the source behind every figure or claim you plan to repeat, check the page says what the report says, and look at who published it and when. When something went wrong along the way, a Limits of this report section says what, for example that some pages couldn't be read or that only one source could be read. Treat a report built on one or two sources as a starting point. On the phone, if a report stops short with no Sources list, reload the chat to see all of it.
Practice. Pick the two claims in a report that matter most to you, open their sources, and write down any claim you can't find on the page.
Before the report goes into Daniel's board paper, Grace Liu, the centre's finance manager, opens source [2] to check the point about tariffs.
When a run fails or comes back thin
Read the failure message, then try again later or reword the question.
When deep research can't finish, the reply says so instead of inventing a report. If the search service didn't respond, the message says no sources were read and no report was written, and that this is a problem on Oppermind's side, not a sign that nothing exists on the topic. Wait a few minutes and send the question again. Do the same when it found pages but couldn't open any of them. If the search worked but nothing usable came back, it asks you to rephrase or broaden the question. If it read pages but couldn't write the report, it lists the sources it read so you can open them yourself. A report that did arrive can still be thin, and Limits of this report tells you why. Run it again with a broader question, or ask about one part in the same chat. A follow-up that isn't worded as a deep research request gets an ordinary chat reply rather than a new run.
Practice. When a run fails, copy the question, wait a few minutes and send it again with Deep research: at the start.
Deep research: What rebates are available in Tasmania for switching a small health centre from gas to electric hot water?
Keep going
Make the learning stick
Start with one real piece of work.
Open Oppermind beside the lesson, apply each step, and leave with something you can use.