Patterns That Help

Once you're being clear about what you want, a few repeatable patterns reliably push the answer from okay to good. None of these are tricks. They're ways of giving the model more to work with. Reach for them when a plain request isn't landing.

Give it a role

Telling the AI who to be sets the tone, vocabulary, and priorities of the answer in one move.

You're an experienced kindergarten teacher. Explain why my 5-year-old
melts down at bedtime, in plain language, like you're talking to a tired
parent.

versus the same question with no role - which tends to come back clinical and hedged. A role isn't roleplay for its own sake; it's a shortcut for "answer the way this kind of person would." "Act as a skeptical editor." "You're a budget-conscious financial planner." "Respond like a patient IT helpdesk person." Each one shifts what the model emphasizes and what it leaves out.

Keep it honest, though. A role makes the style fit; it does not make the model an actual licensed professional. "Act as a doctor" doesn't turn it into one - treat medical, legal, and financial output as a starting point to verify, not advice to follow.

Ask for steps

For anything with reasoning or sequence - planning, troubleshooting, comparing options, math - ask the model to lay out its thinking or work through it in steps rather than blurting a final answer.

Walk me through how to decide whether to repair or replace my 9-year-old
washing machine. Lay out the factors step by step, then give a
recommendation at the end.

This does two things. It usually produces a more careful answer, and it lets you see the reasoning, so when something's off you can spot exactly where it went sideways instead of arguing with a verdict. (Many newer "reasoning" models do some of this internally now, but explicitly asking still helps for everyday tools and everyday questions.)

Show an example (few-shot)

This is the most underused pattern and one of the most effective. If you want output in a particular style or shape, show one or two examples of what "right" looks like. The model is very good at matching a pattern it can see.

Say you're turning rough notes into clean meeting action items:

Turn my messy notes into action items. Match this format exactly:

Example input: "talk to sam about budget, the report is late, maybe move
the launch"
Example output:
- [ ] Sam - discuss budget (owner: me)
- [ ] Chase the late report (owner: ?)
- [ ] Decide whether to move the launch date (owner: me)

Now do these notes:
[your notes]

Showing one example beats three paragraphs describing the format. This trick - giving examples - is often called "few-shot" prompting; giving none is "zero-shot." You don't need the jargon, but you'll see it around.

Pin the output format

Phase one mentioned format; it's worth its own line because it saves the most cleanup. Tell the model the exact shape you want and it'll usually comply:

• "Answer as a table with columns: Option, Cost, Effort, Best for."
• "Give me exactly three options, no preamble."
• "Reply with only the rewritten text, nothing else."
• "Bullet points, no longer than one line each."

"No preamble" and "nothing else" are quietly effective - they cut the "Certainly! Here's..." throat-clearing and the unsolicited summary at the end.

Let it think before it answers

Related to "ask for steps," but broader: for harder requests, tell the model to plan or check before committing.

Before you write the cover letter, first list the 3 things this job posting
seems to care about most. Then write the letter to hit those three.

You can also ask it to review its own work: "Now reread that and flag anything that sounds generic or untrue." It won't catch everything - a model checking itself has real limits and will still miss its own mistakes - but it catches more than zero, and it's free.

Combine them, don't overload

These stack. A genuinely strong prompt might assign a role, give context, show one example, and pin the format - all in a few lines. But don't pile on every pattern for a question that didn't need any of them. "What's a good substitute for buttermilk?" needs none of this. Match the effort to the stakes. The patterns are tools in a drawer, not a checklist you run every time.