How do you plan an AI-assisted research project step by step?

Research tools built on AI can now search dozens of web pages, read PDFs and hand back a report with citations in a few minutes. When OpenAI introduced deep research, for example, it said a task would take five to thirty minutes, with a sidebar showing the sources used. That speed is real, and so is the risk: OpenAI itself warned that the tool can make up facts, can struggle to tell an authoritative source from a rumor, and often fails to say when it is unsure.

A plan solves most of that. Instead of typing a big question and trusting whatever comes back, you split the project into stages, decide in advance what counts as good evidence, and give the AI the parts of the work it does well. The five stages below work for a school assignment, a work report or a personal decision.

Stage 1: The question and the decision

Write down the question in one sentence, then the decision or output it serves. "Electric cars" is a topic. "Would an electric car cost our family less than our gas car over eight years, given about 15,000 km a year?" is a question, and "decide what to buy next spring" is the decision. A precise question tells the AI what to search for and tells you when you are done.

Add your limits: the place (your country or province), the time span (sources from the last three years, say) and the depth (a one-page summary, not a thesis).

Stage 2: The sources that count

Before any AI search, list the kinds of sources you will accept as evidence: official statistics, government agencies, peer-reviewed studies, manufacturers' specifications, reputable news outlets. Also list what you will treat only as leads, such as forums, blogs and social posts. This list becomes part of your prompts, and it is your defense against the rumor problem.

Stage 3: Narrow jobs for the AI

Anthropic, describing how its own research agents work, found that each search task needs a clear objective, an output format, guidance on which sources and tools to use, and limits on scope. Vague assignments led its agents to duplicate work or miss parts of the question. The same applies when you are the one giving instructions.

Here is a worked prompt for the car question:

Research question: Would an electric car cost a family in Quebec less than a gas car over eight years, driving about 15,000 km a year?

Your job in this step is ONLY to gather evidence, not to answer.
- Find the cost factors that matter: purchase price after rebates, energy cost per km, maintenance, insurance, resale value.
- For each factor, give 1–3 facts, each with the source name, the date and a link.
- Use government sites, utility rate pages, consumer organizations and reputable news. Do not use forums or social media as evidence.
- Make a separate list called "Disagreements" where sources give different figures.
- Make a list called "Not found" for anything you could not support with a source.

Then use separate rounds for separate jobs: one to summarize a long report you uploaded, one to build a comparison table from facts you have already checked, one to argue the opposite conclusion so you can see the weak points.

Stage 4: Checks

Keep a small log. For each fact you plan to use, record it with its source, date and whether you opened the source yourself.

The rows are illustrative. The rule is not: if the "Opened" column says no, the fact is not ready. TechCrunch asked, when deep research launched, whether people would really check the citations in long AI reports. This log is how you make sure the answer is yes.

Stage 5: The write-up

Write from the log, not from the AI's report. Ask the AI to draft only from the checked facts you paste in, and to mark any sentence that goes beyond them. Then add a short "limits" paragraph: what you could not find, where sources disagreed, and what would change your conclusion.

Where should your own judgment come in?

At the start, you decide what question matters and what decision it feeds. An AI can suggest angles, but it does not know your budget, your values or your situation unless you tell it, and it cannot tell you which of its angles matters most to you.

In the middle, you judge sources. Deciding that a utility's rate page outweighs a forum post, or that a two-year-old figure is too old, is a judgment about trust, and research tools are known to be weak at it. When sources disagree, read both and decide why they differ; often they measure different things.

At the end, you own the conclusion. Weigh the evidence, state your uncertainty honestly, and put your name only on claims you have checked. The AI did the legwork. The judgment of what is true, and what to do about it, stays with you.