Prompting, the next level

In Everyday prompting you learned the four ingredients of a good request: goal, context, format, and an example. That gets you a solid first answer. This lesson is about what comes next, the moves people who use AI every day reach for when a solid first answer is not enough.

None of these are tricks or secret phrases. They are habits that work because of how the model behaves: it predicts what a good response to your exact words looks like, and it does better when it has something concrete to aim at, room to work things out, and a second pass to catch its own mistakes. Anthropic and OpenAI say much the same in their own published prompting guides, as of October 2026.

Here is the same everyday request, a welcome message for a new hire on a Hurricane landscaping crew, run through each technique. Tap one, then flip between before and after.

Show it, do not describe it

The single most useful upgrade is to give the AI examples of what you want. The fancy name is few-shot prompting: a "shot" is one example, so few-shot just means "a few examples included in the prompt." Zero-shot means you described the task and gave no examples at all, which is how most people prompt.

Anthropic's prompting guide for Claude calls examples "one of the most reliable ways" to steer the format, tone, and structure of an answer, and suggests three to five of them for best results. OpenAI's guide describes the same idea as including "a handful of input/output examples" in the prompt. Both say the same thing about picking them: make them varied. If all three of your examples start with "Hey," every answer will start with "Hey," whether you wanted that or not.

Good examples are real ones. Paste two texts you actually sent, the reply to a customer you were proud of, or a past menu description that sold well. Then say what to do with them: "Match the length and tone of these. Do not copy the wording." That last line matters. Without it, the AI sometimes borrows whole phrases from your samples.

Examples also work for things that are not writing style. Want a spreadsheet of jobs sorted a certain way? Show it two rows done right. Want a summary that always pulls the same three facts out of an insurance quote? Show it one quote and the summary you would have written by hand.

Make it plan, then make it check itself

Two techniques work on the AI's process instead of its answer.

Ask for the plan first. Before it writes the thing, ask it to list the steps, the outline, or the points it plans to cover. "Before you write the email, list the five points it needs to make. Wait for me to approve the list." You get two wins. The model writes each word based on everything before it, so a good outline on the page steers everything that follows. And you get a cheap chance to fix the direction before it spends a page going the wrong way. Fixing a five-line outline is easy. Fixing a finished document is not.

Some apps now have a built-in version of this called a thinking or reasoning mode, where the model writes working notes before it answers. That is covered in Thinking models. Asking for a plan is still worth it with those models, because the plan is something you get to read and change.

Run a critique loop. A critique loop is three steps: get a draft, ask the AI to find the weaknesses in it, then ask it to revise based on its own critique. Anthropic's guide names this draft, review, refine pattern as the most common way to chain prompts together, as of October 2026.

The trick is to give the critique a point of view. "What do you think?" gets polite praise. "Read this as a homeowner in Washington who just got a surprise bill. What would annoy you?" gets real problems. You can also hand it a checklist: "Check this against these three rules and tell me which ones it breaks."

One honest warning: the AI grading its own work is better than no grading, but it is not a second opinion from a person. It can miss the same blind spot twice. The critique loop catches sloppy drafts. It does not replace you reading the final version.

Let it ask the questions

Most bad answers come from missing information, and the person asking usually does not know which information is missing. So flip the conversation. Tell the AI what you want, then ask it to interview you first.

The magic words are something like: "Before you start, ask me whatever you need to know to do this well. Ask one question at a time, and stop after five." The "one at a time" part keeps you from getting a wall of twelve questions you will skim. The "stop after five" part keeps it from interviewing you forever.

This works especially well for things you have never done before. Planning a family reunion at Sand Hollow, writing a rental agreement for a side-by-side, picking a health insurance plan during open enrollment. You do not know what you do not know, but the model has seen thousands of these and knows what usually matters. Let it pull the details out of you, then let it write.

A small variation: when you get an answer that feels off, ask "What did you have to assume to write this?" The list of assumptions often shows exactly where it guessed wrong.

Name the shape, and chain the small jobs

Ask for a format by name. You already know to ask for a length and a tone. Go further and name the shape: a table with specific columns, a checklist with boxes, a numbered list of steps, a two-column comparison, a short text plus a longer email version. For example: "Compare these three quotes for a gutter repair in a table with columns for price, warranty, start date, and anything unusual." A shape you name is a shape you can paste straight into a text, a note, or a spreadsheet without reformatting.

Formats also make the AI more careful. A table with a "warranty" column forces it to find the warranty for every quote, or to leave an obvious blank you will notice. A paragraph lets it quietly skip one.

Break big jobs into a chain. A prompt chain is a series of small prompts where the output of one becomes the input of the next. Instead of "plan my whole open house, write the invites, and make the shopping list," you go step by step: first the guest count and schedule, check it, then the invite using that schedule, check it, then the shopping list from the guest count.

Chains win for the same reason a good contractor does one room at a time. Each step is small enough to check. A mistake in step one gets caught before it spreads into steps two and three. And the model gives each small job its full attention, instead of rushing the last part of a giant request.

If you find yourself typing the same setup at the start of every chat, "I run a tint shop in St. George, keep it short, no emojis," that belongs somewhere more permanent. That is what System Prompting covers: setting instructions once so they stick across every chat.

Try this yourself

Pick one small writing task from your real week: a text to a customer, a note to your kid's coach, a reply to a neighbor. Open the chat app you already use and run this in a new chat:

I need to write: [describe the message in one sentence].

Before you write anything, ask me up to three questions you need answered. Ask them one at a time.

After I answer, write a first draft. Then list the two weakest spots in your draft and rewrite it to fix them.

Then, in a second new chat, ask for the same message with just the one-sentence description and nothing else.

Put the two results side by side and notice:

• Which details did the questions pull out that you would have left out?
• Did the critique find something real, or just nitpick?
• Which version could you actually send without editing?

For bonus points, paste one real message you have sent before into the first chat and say "match this style." Watch how much closer it gets to sounding like you.