I did a live stream walking through prompt engineering end to end, and if I had to keep one move from it, this is the one: show the model an example. Tell it what good looks like instead of describing it, and the quality jumps. Here is the gist, so you get the value without watching the whole thing.
Start with context, then show it
Anthropic’s own guide says to treat Claude like a brilliant but brand-new employee who has no context on your norms or style. In the stream I test this with “write me a poem.” On its own that gives a generic result, so I add what it is for (a comedy club intro), who the audience is, that it is a first draft, and that it should be short and modern. Just that context turned a flat request into three usable, funny options.
Then comes the bigger move: examples. Describing the output leaves room for interpretation, an example removes it. No example is zero-shot, a few is few-shot or multi-shot, and Anthropic calls three to five good samples your secret weapon for accuracy and consistency. When I pasted two popular funny poems and said “generate funny ones like these,” the model picked up the parody and theme and matched the style. The harder or more style-specific the task, the more examples earn their place.
Beyond examples
Once examples are doing their job, a few structures handle the rest. I cover the same toolkit in my first-ever live stream on the channel, embedded here.
XML tags mark clearly where an example ends and your instruction begins, which helps the model parse the prompt. A system prompt sets a role for the whole session, like “you are a seasoned data scientist at a Fortune 500 company,” turning a general assistant into a domain expert. Chain-of-thought is as simple as adding “think step by step,” which research shows lifts quality on complex tasks. Tree-of-thought explores several reasoning paths before picking one, useful when you ideate business or feature ideas.
Two more from the stream stood out. Meta-prompting asks the model to rewrite your prompt first, then you paste that cleaner prompt into a fresh session for a better answer. Reflection asks the model to find flaws in its own response and correct them, and in the demo the corrected answer was noticeably sharper than the first pass.
Keeping it grounded
A prompt-format example only helps if the answer stays truthful. AI will invent metrics if you let it, so the stream shows telling the model plainly not to make things up and to use only the facts you provide, asking you for real numbers rather than guessing. React and RAG extend that idea by letting the model reason, act, and pull in outside information instead of answering blind.
Watch both walkthroughs above for reusable prompt examples, and subscribe for more AI engineering that you can apply the same day.
I specialize in helping teams write prompts and build AI workflows that hold up in production. Book a call at cloudyeti.io/meet.