Prompting Techniques That Level You Up: Examples, Step-by-Step & Building Knowledge

So far in this series we've covered specificity and a checklist of good habits. Now we move to named techniques — repeatable patterns that experienced users reach for when a plain request isn't enough. The good news: these have intimidating names but simple ideas. Master three of them — showing examples, asking for step-by-step reasoning, and building up knowledge first — and you'll handle far more demanding tasks with confidence.

Technique 1: show examples ("shot" prompting)

One of the most reliable ways to steer an AI is to show it what you want instead of only describing it. This is called "shot" prompting, where a "shot" simply means an example. The number of examples you provide gives the technique its name:

Zero, one, and few-shot prompting

Zero-shot

No examples

"Turn this product note into a one-line tagline."

You rely entirely on the model's general ability. Fast, and often fine for simple tasks.

One-shot

One example

"Turn this note into a tagline. Example: 'Noise-cancelling earbuds' → 'Silence, on demand.'"

A single example pins down the style and tone you're after.

Few-shot

A handful of examples

"Turn this note into a tagline. Here are three examples of the voice we use…"

Several examples set firm boundaries, so the output stays consistent.

More examples mean tighter control. Start at zero-shot; add examples when the style or format matters.

The pattern is wonderfully general. Want the model to classify feedback as positive or negative, format addresses a certain way, or match your brand's writing voice? Give it two or three worked examples and it will follow the pattern. Examples communicate nuance that's hard to put into rules.

Reach for examples when tone, format, or style matters more than facts. If you find yourself struggling to describe exactly what "good" looks like, stop describing and just show one.

Technique 2: ask for step-by-step reasoning (chain-of-thought)

For anything involving logic, planning, or multiple moving parts, asking the model to work through it in steps dramatically improves the result. This is called chain-of-thought prompting. Instead of demanding the final answer immediately, you invite the model to reason out loud — and to build the solution one stage at a time, with you steering between stages.

Chain-of-thought: one stage at a time

"Let's plan my budget step by step."Set the goal and invite reasoning

Step 1 — list income sourcesModel lays out the foundation

Step 2 — group monthly expensesYou confirm, then continue

Step 3 — set savings targetsBuilds on the prior steps

A complete, reasoned planEach stage checked along the way

Breaking a problem into ordered steps lets you verify each stage instead of trusting one big leap.

Two things make this work. First, the model produces better conclusions when it reasons before answering, rather than committing to a guess up front. Second, by going stage by stage, you stay in control — if step two looks wrong, you fix it before it pollutes step three. It's the difference between reviewing a plan as it's built and discovering a flaw only at the very end.

Technique 3: build knowledge first (generated knowledge)

Sometimes the best way to get a great answer is to have the model prepare before it produces. The generated-knowledge approach splits the work into two moves: first ask the model to lay out the relevant facts or principles, then ask it to use that material to do the actual task. The first answer becomes high-quality context for the second.

First prompt: "List the key principles of a healthy weekly meal plan, with a sentence on why each matters."
Then: "Using those principles, create a 7-day meal plan for a busy vegetarian."

Because the second request is now grounded in the principles the model just articulated, the plan it produces is more thoughtful and consistent than if you'd asked for it cold. You're essentially letting the model assemble its own reference material — then holding it to that material.

Notice how these techniques layer. Generated knowledge is really chain-of-thought stretched across two prompts, and you can sprinkle in examples (shot prompting) at either stage. They're building blocks, not rival choices — combine them to fit the task.

Choosing the right technique

A quick rule of thumb for when to reach for each:

Show examples when style, tone, or format is what matters — classification, rewriting, matching a voice.
Ask for steps when the task involves logic, planning, or several interdependent parts.
Build knowledge first when you want a richer, more grounded result on an open-ended or expert topic.

And when in doubt, combine them: a few-shot example plus a request to reason step by step is a powerful default for tricky work.

The takeaway

Techniques sound advanced, but each rests on an everyday instinct: show what you mean, think before you answer, and gather your facts before you start. Shot prompting, chain-of-thought, and generated knowledge give you reliable patterns for the moments when a plain request falls short. In the final article of the series, we'll look at context — how much an AI can keep in mind at once, and how the right background quietly shapes everything it says.

Dhaval Shah

Cloud Solution Architect — 18+ years across SharePoint, Microsoft 365, and Azure AI.

Back to Blog