Asking better questions
You have already met the principle (Context is king) and the four-part frame from Chat agents. This course is the next layer down: what to do when the first answer comes back generic, and how to get unstuck without giving up on the session.
The first answer is rarely the best
A generic answer is almost always a context problem, not a model problem. When the first reply is too vague, too obvious, or just slightly off, the right move is not to retry the same prompt or switch to a more capable model. It is to add what was missing. The model is doing its best to read your mind; the fix is to stop making it guess.
The pattern is simple: you write a prompt, the answer is not great, and instead of typing a follow-up complaint (“be more specific”), you write a sharper version of the original. Pasting the previous bad answer back in works too: “the version below is too generic, here is what is missing”.
Five moves that push a generic answer into a useful one
Five small adjustments cover most of the cases where the first answer falls flat.
- Add the constraints. “in three sentences”, “for a senior engineer”, “without using the word X”, “by Friday”. Constraints turn a wide question into a narrow one.
- Add the audience. “Explain this to a junior PM” lands very differently from “explain this to me”. The audience is half the prompt.
- Add what you tried. “I have already considered X and Y, here is why neither worked.” Moves the conversation past the obvious answers in one turn.
- Ask for options instead of an answer. “Give me three different ways this could go, with the trade-offs.” Lets you pick instead of accepting whatever the model produced first.
- Show, do not tell. Paste an example of what good looks like. “Here is a status update I wrote last week and liked. Match that tone.”
When to switch tools, when to stay
Most “the model is bad at this” complaints turn out to be a question problem. Before you switch tools, check whether you tried any of the five moves above. If you did and it still falls apart, then a different agent might fit better: Claude for long careful work, Gemini for data and Workspace context, NotebookLM when you need everything grounded in your own sources. But the order matters: tighten the question first, switch tools second.
A useful test: rewrite the prompt as if you were paying a colleague to do the work. If the rewritten prompt would let a real human do a good job, the agent will probably handle it. If a colleague would still be confused, switching tools will not unconfuse the model either.
Hands-on
Open the chat history of whichever agent you use most. Find a prompt from the last week where the answer was not great. Without changing tools, rewrite the prompt by adding two of: constraints, audience, what you tried, “give me three options”, an example. Re-ask.
Pick a question you want a real answer to today. Before sending, ask the agent: “what context do you need from me to give a good answer to this?”. Read what it asks for, paste in the missing pieces, and send the real prompt only after.
Take a recent ambiguous brief from your work (a Slack thread, an email, a doc). Ask the agent to “list five different ways I could interpret this brief, with the trade-offs of each”. Pick the one that matches what you actually meant, and use that as the basis for the next question.
Reflect
- Which of the five moves do you already do without thinking, and which feels least natural? The one you avoid is usually the one that would help most.
- When you get a generic answer from the model, what is your first instinct: switch tools, retry the prompt, or rewrite the question? The third one is the right one. How often do you actually do it?
View as plain markdown for LLMs and copy-paste