Business workflow guide
ChatGPT dots research: build a brief you can check
AI-assisted guidance prepared for Blue Canvas from the official sources linked below. The examples are proposed workflows, not reports of firsthand product testing.
A research brief should help someone make a decision without repeating the whole search. Ask for a small set of supported findings, a visible list of unknowns and links to the evidence behind each important statement. More pages do not necessarily make a better brief.
Start with the decision, not a broad topic
OpenAI describes research as one of the things a dot can do. The workflow below is our suggested way to commission and check that work, not a dedicated research mode or a promise of complete coverage.
Replace 'research event venues' with a question such as 'which of these three venues should we ask about for a small workshop?' State the location, essential facilities and what would rule an option out. Keep a shortlist separate from a booking decision. For an initial exercise, use public pages and fictional requirements rather than attendee details or private negotiations.
Agree when the evidence must be current. A venue's contact page might establish its address, but an old brochure cannot confirm availability for your date. Ask for the date of each relevant statement, where provided, and the date it was accessed. Missing dates belong in the limitations.
Product source: OpenAI's official documentation, checked 31 October 2026.
Make every important claim traceable
Choose sources that can support the particular question. A venue's own facilities page is a sensible starting point for a room specification. A search-result summary is only a lead to inspect. If the page cannot be opened, the brief should say so instead of presenting its snippet as verified evidence.
Use a claim table rather than a bibliography detached from the conclusions. A reader should be able to find the original passage without guessing which link supports which sentence. Where sources disagree, keep both positions visible and explain what remains unresolved.
| Field | What to record | Why it matters |
|---|---|---|
| Finding | One specific statement | Avoids combining several unsupported claims |
| Evidence | Direct page link and relevant heading | Lets the reviewer inspect the original |
| Scope and date | Which location, service or version; checked date | Stops an old or partial answer becoming general |
| Uncertainty | Missing detail or conflicting evidence | Shows what still needs a decision |
Keep a factual finding separate from a recommendation
For a fictional workshop, 'the page lists two meeting rooms' is a finding if the source says that. 'This is the best venue for us' is a judgement requiring criteria the page cannot supply. Ask the brief to label recommendations and explain which agreed requirements they satisfy.
Treat unanswered requirements consistently. If one venue does not publish access information, mark that field unconfirmed; do not rank it as accessible or inaccessible. A shorter shortlist with clear gaps can be more useful than a confident ranking built on assumptions.
Keep quotations short and necessary. Usually a concise paraphrase and a direct link are enough. Do not ask for whole articles to be copied into the pack. The useful output is your decision aid, not a collection of other people's pages.
Use a bounded research brief
Supply the question, permitted sources and stopping point together. If the initial sources are inadequate, the useful result is an explanation of the gap and a proposed next research step. It is not permission to create accounts, contact providers or buy access.
Run this as a one-off exercise first. Review the evidence table before asking for a polished executive summary: that makes it easier to correct weak foundations before they disappear inside fluent prose.
Adapt this example to your approved task
Prepare a one-off research brief for [decision] using only [approved public sources]. Apply [essential criteria] and stop after [bounded scope]. For each finding, give the direct source, relevant heading, source date if available and date checked. Separate facts, inferences and unknowns. Do not treat inaccessible pages or search snippets as verified evidence. Return the evidence table and a short recommendation for review in this conversation. Do not contact anyone, sign up, purchase, publish or create recurring work.
Review the result before expanding the responsibility
OpenAI distinguishes assigned tasks from proactive research: the latter can examine permitted information and keep notes, but does not itself take actions in connected apps. Do not assume that a background suggestion has answered the specific research brief you agreed.
Open every source supporting a decision-critical claim. Check that the text actually supports the claim, that the relevant option is the same one you are considering, and that a limitation has not vanished from the summary. Recheck volatile details when the decision is made, not only when the brief was prepared.
Finish with three lists: what is established, what needs a human answer, and what was outside scope. If the first result is unreliable, narrow the question or improve the source list before commissioning more. A useful pilot can end with a decision to keep this task manual.
Product source: OpenAI's official documentation, checked 31 October 2026.