4 Prompt Engineering Tips That Changed How I Use AI

aiprompt engineeringproductivity

Most prompt engineering tips tell you to be more specific. Add more steps, more rules, more examples. I followed that advice for a long time, and it made my AI tools worse at the exact things I wanted help with.

This is what I used to do. I took a task I already knew how to do, wrote down every step and taught the AI to follow it. Write the email like this. Research in this order. Use these three sources. It felt responsible. It felt like good prompting.

Then I read a breakdown of an interview with Felix Rieseberg, who leads engineering for Claude.ai, Cowork and Claude Code Desktop at Anthropic (watch the interview). It made me rethink my whole AI setup. These are the four prompt engineering tips I took from it, and how I use each one at work.

1. Describe the result and let AI find the route

Tell the AI the result you want and how you'd know it's good, then let it work out the path. That's the whole tip. The reason it works is that your own workflow carries your limits.

Every workflow you've built reflects three things: what you knew how to do, the tools you'd learned, and how much time you had the day you set it up. All three are just where you were at the time.

When you make AI follow that workflow step by step, you hand it the same limits. It can't use a tool you've never heard of, because you didn't list it. It can't skip a step that turned out to be pointless, because you told it to do every step. It can't take a faster route, because you already picked the route.

That's the catch with a lot of AI prompt best practices. They push you to control the process, and the process is often the weakest part of the whole thing.

What changed when my AI chief of staff got goals

To judge the result, you only need a clear idea of what good looks like.

I learned this with my own AI chief of staff. For the first couple of weeks I gave it tasks and rules. It sorted my messy brain dumps, planned my days and broke scary tasks into small steps. It kept me very busy, and neither of us knew if any of it was going anywhere.

Then I gave it my 12-week goals and the two numbers I check every week: 3 sales moves and 10 videos shipped. Now anything that doesn't serve those goals goes to the ideas inbox. My side projects are not loving it. Nothing else changed. The AI finally knew what good looked like.

This is also the idea behind context engineering, the term you'll see everywhere in AI right now. The goal, the quality bar and the background you give the AI matter more than a perfect list of instructions.

How to use AI at work with result-first prompts

A result-first prompt answers five questions. Plain, honest answers are enough.

  1. What is the finished thing? A shortlist, a draft email, a one-page summary, a decision.
  2. Who is it for and what will they do with it? Your boss deciding on budget reads differently from a customer reading an offer.
  3. How will you judge it? This is the easiest part to skip, and it does the most work. Write down what would make you say "yes, that's it".
  4. What are the hard limits? Budget, deadline, tone, things it must never do.
  5. What should it do when it's unsure? Ask you, flag it, or make its best guess and say so.

Here's the same task written both ways. Say you run sales for a small B2B company and need a new CRM.

Before (the workflow prompt):

Go to the websites of these five CRM tools. Find the pricing page for each. Copy the price of the mid-tier plan into a table. Then check the review sites for each one and write down the star rating. Then write a summary.

After (the result prompt):

I need to pick a CRM for a 6-person B2B sales team by Friday. Give me a shortlist of 3 I could realistically choose between. A good answer tells me the real monthly cost for 6 users, what each one is bad at, and which one you'd pick for a team that lives in Gmail. If you can't verify a price, say so instead of guessing.

The first prompt gets you a table of five star ratings. The second one leaves the AI room to find a tool you didn't list, notice that the "mid-tier" price is only for annual billing, and dig into how well each one syncs with Gmail. You told it what matters, so it can go looking for it.

2. Skip "act as an expert" and ask for the expert's methods

"You are a top copywriter" might be the most copied prompt trick on the internet. It even has a name, role prompting, and I used it for ages. The problem is that the AI picks up how a copywriter sounds, and the thinking underneath stays the same. You get confident words on top of average work.

What works better is asking for the things a professional actually does with the task. A good editor pushes on your argument, hunts for the weakest claim and argues back.

So instead of "Act as a senior marketer and review my landing page", I ask:

  • What would a skeptical buyer object to on this page?
  • Which claim is the weakest, and what proof would fix it?
  • What's the strongest case for not buying this?

Each question is a real step a pro would take. If you don't know what those steps are in some field, ask the AI first: "What does a professional do when they review a landing page?" Then have it do exactly those things.

3. Remove instructions before you add more

When a draft disappoints, the reflex is to add a rule, then another one, then a checklist. Give it a few months and you have a very detailed setup, and you're still fixing drafts.

A lot of those rules were written to patch weaknesses that older models had. Models improve every few months, and old patches can start getting in the way. Sometimes the best fix is deleting instructions.

The test takes about ten minutes:

  1. Pick a task you repeat, like a weekly LinkedIn post or a client update.
  2. Run it with your usual instructions, custom prompts or skills.
  3. Run the same task with the same brief in a fresh chat, without them.
  4. Check which version needed less fixing. Keep only the rules that still earn their place.

When steps still matter

Starting with the result still leaves room for instructions. Steps earn their place in a few cases:

  • A fixed format. If the output goes into another system, a template or a report your CFO expects, spell the format out.
  • Rules you can't bend. Legal wording, compliance checks, things that must happen in a certain order.
  • A known weak spot. If the AI keeps skipping something, like double-checking the numbers, add that one step.

My rule of thumb: start with the result, run it, and only add a step when the output keeps going wrong in the same way. Most of the time you'll add one or two steps, and you'll keep the AI's better route for everything else. For me, that's how to use AI effectively without babysitting every step.

4. Review the work in a fresh chat

Asking the same chat "is this good?" right after it wrote the thing is like asking someone to grade their own homework. It remembers why it made every choice, so it tends to go easy on them.

Open a new chat for the review. Paste in the finished draft and the quality bar you wrote in tip 1, and ask what's weak. A fresh session has no reason to protect the draft.

This is where the four tips connect. The result and quality bar from tip 1 become the checklist the fresh chat grades against, so you write them once and use them twice.

FAQ

What is the most useful prompt engineering tip?

Tell the AI what result you want and how you'll judge it, before you tell it how to work. A clear quality bar beats a long list of steps for most everyday tasks.

How do I prompt AI to get better results?

Describe the finished thing, who it's for, what makes it good, your hard limits, and what to do when it's unsure. Then compare the output to your quality bar and adjust the prompt from there.

Is prompt engineering a hard skill?

No. It's closer to briefing a smart new colleague than to writing code. If you can explain what a great result looks like, you can write a good prompt.

What is the difference between prompt engineering and context engineering?

Prompt engineering is about how you word a single request. Context engineering is about everything the AI knows when it answers: your goals, background, examples and quality bar. Result-first prompting is a simple way to start doing both.

Does role prompting still work?

It changes the tone more than the quality. Asking for the specific steps a professional would take gets you better work than telling the AI who to be.

Should AI review its own work?

Use a separate, fresh chat for reviews. The chat that wrote the draft tends to go easy on it.

Try it this week

Pick one prompt you use every week. Delete the steps. Write down what a great result looks like and how you'd check it. Run both versions side by side and see which one wins.

My guess is you'll find out the same thing I did: a good chunk of my "process" was just habit.

If you want to try this on your own team's real tasks, it's exactly what we practise in my AI workshops for business leaders.