How to hand an AI your research materials

Most disappointing AI research comes from one of two mistakes: the AI didn't have the material it needed, or it wasn't clear what the result was for. Both are fixable before you press send.

Start with the decision, not the topic

"Research urban heat" produces a survey. "Research how urban tree cover affects summer heat, explain what the evidence supports and where cities differ, and use original research and public-agency sources where possible" produces something you can use.

A good research brief has three parts:

• The question, stated as something that can be answered.
• The scope: time period, geography, what's out of bounds.
• The result: a recommendation, a comparison table, a one-page brief, a list of open questions.

Give it the sources you already trust

If you have the report, the contract or the dataset, don't make the AI find a worse version of it on the web.

In Manwe there are two ways to do this, and the difference matters:

• Attach a file by dragging it into the composer. Manwe keeps a copy of that file for the conversation. It doesn't get access to the folder it came from.
• Choose a working folder for ongoing work. The folder panel decides whether it's read-only or can be edited under the chat's approval mode.

For a codebase, choose the repository folder. Repository search can trace a feature without loading everything into the conversation.

Put the context where it belongs

If the research is part of ongoing work, select the right World first. A Deep Review run inside a World works from that World's goals, constraints and known facts. Your budget, your audience and your deadlines shape the conclusion, not just the general literature.

If a finding should outlast the conversation, make sure it ends up in memory, or in Review if it came from a document and you want to approve it first.

Ask for the shape of the evidence

Ask the AI to tell you not just what it concluded, but how sure it is and why:

• "Separate what's well established from what's contested."
• "Say what would change your conclusion."
• "Flag any claim that rests on a single source."

Deep Review is built around this. It reports the conclusion alongside the supporting evidence, disagreements, risks and open questions.

Check what you'll rely on

The final step is still yours. Open the sources behind any claim you plan to act on. A long list of citations isn't the same as good evidence, and a confident tone isn't proof. If a source can't be opened, a good report says what that changes. You should also know which conclusions survived without it.

AI makes the reading fast. It doesn't make the judgement unnecessary. The best research setups are the ones that make the judgement easy to exercise.