10 Useful AI Prompts for Research, Writing and Thinking About Superintelligence

Most weak AI answers trace back to a weak request. A bare question gives the model room to guess your goal, invent missing facts and answer in a shape you cannot easily check. The ten prompts below close that room. Each one is a template you can paste, fill in the brackets and run.

What do all ten prompts have in common?

• Material first. They tell the model what to work from, so it does not fill gaps from memory.
• A fixed output shape. A table, a list or a labelled section is easier to scan and to check than a paragraph.
• Permission to be unsure. Each asks for 'not stated' or 'I don't know' instead of a confident guess.
• A check you can run. The output makes it easy to find the claim and verify it elsewhere.

1. How do I get an explanation I can trust to be pitched right?

Why it is built this way: naming what you already know prevents both a patronising and an over-technical answer. The 'least certain' list turns the model's own doubt into your reading list. Limit: a model can be wrong about where it is uncertain, so treat the list as a starting point, not a guarantee.

2. How do I check a claim for the kind of evidence behind it?

Why it is built this way: it separates 'how is this supported' from 'is this true', which is the distinction behind our evidence labels. Banning outside facts keeps the model from quietly correcting or confirming the text from memory. Limit: it grades the passage's sourcing, not the world. A well-cited claim can still be wrong.

3. How do I summarise a document without the model adding things?

Why it is built this way: requiring a quote makes unsupported statements visible, because there is nothing to quote. Limit: models can still misquote or stretch a sentence, so confirm the quote appears in your source.

4. How do I stress-test a plan before committing to it?

Why it is built this way: starting from failure (a 'pre-mortem') makes it easier to list problems than asking whether the plan is good, which invites reassurance. Limit: the risks are generic unless you give real constraints, such as budget, team size and deadline.

5. How do I get the strongest case against my own view?

Why it is built this way: asking for a fair restatement first stops the model attacking a caricature. Many assistants lean towards agreeing with the user, and this structure pushes against that. Limit: it can still be agreeable. If the pushback feels soft, say so and ask again.

6. How do I get a code review without an unwanted rewrite?

Why it is built this way: forbidding a rewrite keeps the change small enough to review, and the certain/possible tag tells you where to look first. Limit: a model cannot run your code here. Run the tests, and read our guide to choosing a coding assistant for how to judge tools fairly.

7. How do I compare options without a made-up winner?

Why it is built this way: the 'unknown' option stops the model from inventing prices or features, and withholding a winner keeps the decision with you. Limit: prices and plan limits change often, so verify every figure on the vendor's own page.

8. How do I tighten my writing and keep my meaning?

Why it is built this way: the change list makes silent edits visible, which matters when a rewrite can shift a claim's strength. Limit: it can still soften or sharpen a statement. Compare the claims, not just the style.

9. How do I break down a big superintelligence claim?

Why it is built this way: claims about superintelligence often bundle a definition, a forecast and a value judgement into one sentence. Separating them shows which parts can be tested at all. Our guide What is AI? covers the basic vocabulary. Limit: the model may state its own forecasts as if settled. Ask it to label which parts are speculation.

10. How do I make the assistant ask before it answers?

Why it is built this way: many bad answers come from an underspecified request, and a short interview fixes that cheaply. Limit: some assistants ignore the instruction to wait. If yours does, repeat it once.

How do I find out whether a prompt actually works for me?

What are the limitations of this list?

These prompts are our editorial judgement. We have not measured them against alternatives, run them across models, or published results. A prompt that works well today may work differently after a model update. No prompt removes the need to check important facts yourself.