00
How it works / the difference

A good prompt gets you a plausible answer. A method gets you a sound one.

A capable one-off prompt already does a lot. Ask an AI to assess a vendor's claim of 40 enterprise customers, and tell it to check each named customer against their own filings and flag anything it cannot confirm, and it will do just that. It confirms one deployment, fails to confirm another, and marks the headline figure unsupported. Good prompting produces good work.

The catch is that it can only do what you thought to ask for. And the most valuable parts of research are the ones that decide what to look at in the first place. You cannot ask for those, because asking would mean you already had the answer.

01
What a request cannot reach

You set the topic and handed it over. So the AI will not stop and tell you the question worth researching is slightly different from the one you asked.

You can only name leads you already know about. So the small detail that quietly changes the answer, the one you did not know was there, gets missed.

You would have to have already spotted the useful comparison for the AI to lay two things side by side and show you what it reveals.

A standing set of research methods does the whole process every time, including the steps that turn up the questions you did not know to ask. These are things a good researcher does by habit. Written down as instructions the AI follows, they become methods it runs on your behalf.

Work you did not ask for gets done Quality holds across tasks The output can be checked, not just trusted
02
The model / four steps and a base

Four steps, one running base

Research runs through four steps, each resting on the one before, with a set of standards that hold the whole way through. The steps are not strictly one-way: something found late can send you back to redo an earlier part.

03
Bedrock / holds throughout

Standards that run under every step

ST-1Rank the sourcesPrimary and institutional sources support a fact; weak sources only point to leads.
ST-2Match effort to stakesBackground gets one source; a load-bearing claim gets the full chase and the contrary search.
ST-3Keep a running recordEvery claim that matters, its status, its best source, what was searched.
ST-4RevisitA finding that undoes earlier work sends the task back to the step it affects.
04
What it does not do / honest limits

It raises the floor, not to certainty. The method makes a sound result far more likely. It does not guarantee one. The same instruction will occasionally miss a lead it usually catches, which is why it is worth checking a prompt once whenever you change it.

It costs more. The full method might run ten searches on a task a bare request would answer in two. Matching effort to stakes is what keeps that work aimed at the claims that actually carry the conclusion.