LearnGrokMost prompting advice ages badly because it is really advice about one model's quirks. What follows is the part that has stayed true across model generations, in every assistant, including Grok.
1. Supply the context you are silently assuming
The single largest quality gain available to most people is paste. The model cannot see your codebase, your inbox, your document, or the last conversation you had. If the answer depends on something, put that something in the prompt.
A prompt that mentions "our usual format" produces a guess. A prompt that includes an example of the usual format produces the format.
2. Say what the output is for
"Write a summary" and "write a summary I can paste into a status update for people who missed the meeting" produce different lengths, different registers and different levels of detail. The audience and the destination do more work than any adjective.
3. Constrain the shape: not the thinking
Specify the shape of the answer: a table with these columns, at most five bullets, JSON with these keys, a diff. Do not micromanage the reasoning. Asking for a particular format is reliable. Asking a model to "think step by step" is much less load-bearing than it used to be, because current models already allocate reasoning effort themselves.
4. One or two examples beat a paragraph of description
If you can show two examples of input and desired output, do that instead of describing the rules. This is the cheapest form of specification there is, and it is more robust than instructions, which get partially followed.
5. Iterate on the prompt: not the conversation
When an answer is wrong, resist the urge to reply "no, not like that" five times. A long thread of corrections carries all the earlier bad attempts along with it. Edit the original prompt to include what you learned and start fresh; the result is usually better and shorter.
6. Decide up front how you will check it
Before you send the prompt, know what would tell you the answer is wrong. If you cannot answer that, you are not going to be able to use the output responsibly, and no amount of prompt polish fixes it.
What ages badly
- Long "you are a world-class expert" preambles. Occasionally worth a little. Usually noise.
- Threats, bribes and all-caps insistence. Fashionable, unreliable, and embarrassing in a shared prompt library.
- Anything that depends on a specific model version's behaviour. Write prompts you would be happy to keep after the next release.
What changes
How much reasoning a model does on its own, how long an input it accepts, and how strictly it follows format instructions all shift between versions. Re-test your important prompts after a model update, and check docs.x.ai for the current capabilities of whichever model you are calling.
