Practical playbook
Research a topic with reliable sources
You need more than a quick answer, and you want to see which claim came from which source rather than take a tidy summary on trust.
Turn a broad question into a bounded, source-linked briefing you can open up, check and argue with.
- 20 min
- 5 steps
- Comfortable
Use this page in order
Copy the prompt, follow the steps and check the result before using it.
Try it now
Copy this into your AI tool and fill in the brackets.
Remove anything you cannot share first.
Starter instruction
Research [topic] to help me decide [decision]. Focus on [country/industry] and information from [time period]. Prioritise primary sources and credible independent analysis. Compare the strongest options, disagreements and unknowns. Link every material claim. Finish with what the evidence supports, what it does not support and five questions I should investigate next.
Do it
Tick each step as you go.
The person and AI shape the work at different moments. Your ticks stay in this browser.
0 of 5 done
See it done
Watch the judgement change the draft.
Before · the honest starting point
Research heat-pump options for a small Melbourne office before a lease renewal. Prioritise official efficiency guidance and installer-independent sources.
After · the version a person can stand behind
A two-page decision briefing with verified official guidance, the building-specific unknowns named plainly, and five questions to put to qualified installers.
How it got thereAI first pass, then your correction
AI · a useful first pass
Briefing: three system categories, likely building constraints, links to government guidance, and a short list of assumptions about floor area and existing electrical capacity.
You · context changes the answer
The reviewer opens each citation, drops one that was written for a different climate zone, adds the building's real electrical limit, and marks installation cost as unknown until quotes come in.
The decision stays human
Someone must be able to stand behind the result.
The person decides whether the evidence is credible, relevant and sufficient for the real decision. AI can assemble and challenge a case, but it cannot set the organisation's evidence threshold or accept the consequence of being wrong.
Verification required
0 of 4 checked
Go deeper
The detail behind each step.
Open only what you need.
How AI and you split the work3 capabilities
Use the capability. Keep the part that needs a person.
Research becomes useful when it changes the quality of a decision. A long report is not evidence of good research unless the scope, sources, disagreements and remaining uncertainty are visible.
Choose a capability for a real job.
Editable research planning
Deep research features can turn a broad question into a proposed plan, source strategy and sequence of subquestions that the person can review before a long search begins.
- The person contributes
- The person corrects the decision being supported, the geography, time period, definitions and source priorities before the plan hardens around the wrong question.
- Do not assume
- A comprehensive plan can still be precisely aimed at the wrong decision, especially when the original request contains hidden assumptions.
Web, file and connected source retrieval
Research can combine public web material with selected files or connected services such as Drive, Gmail or approved internal repositories, allowing external claims to be tested against organisational context.
- The person contributes
- The person chooses which systems may be searched, distinguishes authoritative records from convenient context, and protects information that should not be included.
- Do not assume
- Access does not equal authority. A connected file may be outdated, informal or written for another purpose, while important paywalled or offline evidence may remain invisible.
A summary with links to its sources
AI can group claims, compare sources, surface disagreement and produce a report with links so the reader can move from summary back to evidence.
- The person contributes
- The person opens the load bearing sources, decides whether the evidence applies locally, and separates a source statement from the tool's inference about it.
- Do not assume
- A citation can exist and still fail to support the sentence beside it, and multiple tools may repeat the same weak source rather than provide independent confirmation.
What you bring that AI cannot3 resources
The resources already inside the work.
Decision framing
You know what action the research must support, which uncertainty is blocking progress, and what level of confidence is sufficient for the consequence involved.
Source and domain judgement
You can recognise primary evidence, conflicts of interest, outdated guidance and definitions that sound similar but mean something different in your field.
Local applicability
You know whether published evidence fits your organisation, population, jurisdiction, systems, budget and operating reality.
What goes wrong, and the skill you buildFailure to watch
Failure to watch
The main failure is false confidence through volume. A long, cited report can hide a poor research question, repeated secondary sources, missing evidence and conclusions that travel further than the sources allow.
Skill strengthened
Evidence judgement
Used well, the workflow strengthens question framing, source evaluation, uncertainty handling and the discipline of matching the strength of a conclusion to the strength of its evidence.
When this is the wrong fit4 signs
- When you need one current fact, go straight to the authoritative source instead of commissioning a whole briefing.
- Do not trust a citation until you have opened it, because the longer the report runs the easier it is for an invented source or a link that does not back the claim to slip through.
- Treat one of these reports as a fast parse of what is already published, not new insight: it skews toward whatever is easy to find, so anything paywalled or off the beaten path simply never makes it in.
- Watch for a briefing that agrees with the answer you clearly wanted, and remember that two tools agreeing is weak comfort when they were trained on much the same material.
Tools that can support this4 tools
Evidence backed starting points, not universal winners. Open a guide or ask Bit to help you weigh the fit.
Its reviewed guide includes Deep Research, which works for several minutes across many sources and returns a report with citations you can click through.
Our read, not tested by usSee why it may fit →GeminiIts reviewed guide includes research that can pull together web material and selected Google files in one pass.
Our read, not tested by usSee why it may fit →KimiIts reviewed guide is built for large document sets and multi-step research that runs over many files.
Our read, not tested by usSee why it may fit →GrokIts reviewed guide includes Search across the live web and X with citations, and Multi-agent, which splits a large question into separate researchers and keeps their source links and disagreements visible.
Our read, not tested by usSee why it may fit →Keep going4 links
Where this came fromReviewed 19 September 2026
- Reviewed
- Review by
- Owner
- The Human Bit editorial
One quick check