Most professionals learn AI by trying prompts and keeping what seems to work. That gets you started, but it does not give you results you can repeat, check, or share with a team. A better path is to learn on one real project, with a goal, a way to measure it, and a few weeks of deliberate iteration. This is a four-week plan you can follow on your own, regardless of the tool you use.
Before you start: pick one project
Choose a piece of work that you do regularly, that is valuable, and that is safe to practice on. Examples: summarizing and classifying customer emails, preparing a recurring report, drafting proposals from a template, or extracting data from documents. Avoid anything that requires confidential or personal data until you know how to handle it. The project is the whole point: you learn the method by building something you will actually use.
Week 1: define the goal and how you will judge it
Anthropic's prompt engineering guide assumes you start with a clear definition of success criteria and some way to test against them empirically. OpenAI's guide likewise recommends building tests and evaluation suites that measure prompt behavior. So in week one:
- Write the goal in one sentence. What should the result achieve?
- Define success criteria. What would make an output acceptable? Be specific about accuracy, tone, format, and what must never happen.
- Collect 10 to 20 real examples. Use real inputs from your work, including a few difficult cases.
- Record your baseline. How do you do this today, and how long does it take?
Week 2: build a first version
Now write your first prompt or workflow. Focus on the basics that vendor guides commonly recommend:
- Give clear instructions. Say what you want, for whom, and in what format.
- Provide reference material. Include the context the model needs instead of expecting it to guess.
- Show examples. A few good input and output examples often clarify what you mean.
- Split complex tasks. Break the work into steps instead of asking for everything at once.
Run your examples through the first version and keep every result.
Week 3: test and refine
This is where professionals separate from casual users. Compare outputs against your success criteria. Look for patterns in the failures, change one thing at a time, and run your examples again to see whether the change helped. OpenAI's guidance recommends building tests and evaluation suites that measure how a prompt behaves, and pinning a specific model version when you need consistent behavior. Keep your prompts in a document or a repository so you can see what changed.
Week 4: make it repeatable
Turn what works into something you or your team can run without you. Write down the instructions, the inputs, the checks, and where a person must review. Decide what to automate and what to keep manual: automate steps that are frequent and low risk, and keep review in steps where an error is costly. Then measure again against your week-one baseline.
Common mistakes
- Changing many things at once, so you cannot tell what helped.
- Judging results by a single impressive output instead of a set of examples.
- Skipping the baseline, so you cannot show improvement.
- Putting confidential or personal data into tools without a clear policy.
- Automating before you have checked quality.
What you should have after four weeks
A working version of your project, a written method, a set of examples you can retest, and a clear view of what to automate next. You will also be able to explain why it works, which matters when others rely on it.
If you prefer to do this with guidance, our 1-to-1 AI training follows a similar structure: four sessions over one month around a single goal that we define in a discovery meeting. For the thinking behind a repeatable method, read From Prompts to Method.