Come with a question, find your answer, get back to work. You navigate by the question you're carrying rather than reading front to back, and any section can read itself aloud.
In the real version, this bucket answers the question that stops people before they start: what information can go into which tool. It would look something like this:
| Kind of information | Approved tool | Not allowed |
|---|---|---|
| Public / non-sensitive | Your org's general assistant | — |
| Internal, not for public release | Your org's protected assistant | Consumer chatbots |
| Regulated / restricted | Your org's secure environment | Everything else |
Plus: what you can and can't put in each tool, human-subjects considerations, and client-agreement limits.
In the real version, this bucket helps a researcher pick a tool and an approach from the task in front of them: a one-screen decision aid, a plain overview of what each approved tool is good at, and a handful of worked examples for the routine chores researchers said they wanted most. Those examples carry organization-specific detail, so they live in the real guide, not this demo.
The first two sections handle permission and tool choice. This one is about judgment: the small set of habits and mental models that decide whether you get real work out of these tools or something that only looks right. Most people rate their own skill here higher than it is, so it's worth a read even if you've been using the tools for a while.
It's built as several short pieces on purpose. Each one is a single idea you can read in under a minute and use the same day, and none of them leans on the others. Read straight through, or jump to whichever piece matches where you are.
The fastest way to get useful work out of one of these tools is to treat it like a new intern. It has read an enormous amount and can connect ideas across fields, and it also has no experience of your project, your standards, or what you're actually trying to do. So it needs what any new hire needs: a brief. Tell it why you need the thing, what a good version looks like, and the background it can't know on its own. You would never hand a new intern a one-line assignment and expect the finished product back. The same holds here.
Whether you can trust what it gives you comes down to one thing: how far it had to reach beyond what you handed it. Give it a document and ask it to summarize that document, and you're in safe territory, because everything it needs is sitting in front of it. Ask it to explain a general concept, and it's usually fine, because that's well-worn ground. Ask it to go find a fact it wasn't given, or to be an oracle about something obscure, and the ground gets soft. Now it's reaching, and when it reaches it will sometimes make things up with total confidence.
The rule of thumb is simple. The more an answer depends on what you provided, the more you can trust it. The more it depends on what the tool went and found, the more you check.
A few tasks look like the tool should handle them, and it quietly won't. Worth knowing these cold. It is bad at arithmetic and counting, because it works in fragments of words rather than numbers, so it gets most of a column right and one cell wrong. It has no reliable sense of today's date or how long a task will take. It doesn't know about events past its training, and it doesn't know its own cutoff. And it will produce citations that look perfect and point to nothing real. Keep this short list handy, and for anything on it, either do the job yourself or check the result every time.
After watching a lot of people use these tools well and badly, the strong ones share six habits. They know what to ask it and what to keep away from it. They know their own role and the tool's role. When they reach for it, it delivers, because they've built the judgment for when it's the right tool for the job. They know the features and actually use them. They work in back-and-forth rather than expecting one perfect answer. And they don't use it like a search engine.
Read that as a self-assessment. The habits you already have, and the one or two worth building next.
Most weak output traces back to a weak prompt, and a solid prompt has four parts you can remember as AGCA. Audience: who is this for. Goal: what does a good result look like. Context: the facts, constraints, and background the tool has no way to know. Action: the actual verb, whether that's write, summarize, compare, or critique.
Before a heavier piece of writing, spend thirty seconds on Step 0. Who is reading this, what do they already know, what posture are you taking, and what should they think or do afterward. Answer those four and the tool stops guessing at the shape of what you want.