Deep dive · module 11 of 100
Refine your prompts
Work with examples, criteria, and targeted improvement steps.
After this lesson
You test a prompt with a normal situation and a difficult edge case.
In view · An improvement loop
Refining without changing everything
Place the answer next to your goal and identify én concrete deficiency.
Request, for example, a shorter text, with the same facts.
Check whether the change helps and does not add new errors.
Part 1 of 3 · View the first version
Your preference for less movement is on. Choose each part yourself.
Show what you mean
A few carefully chosen examples can clarify tone and format. Do not use sensitive data. Also show a difficult case, such as missing information, so the model knows that stating uncertainty is part of a good answer.
Ask for verifiable explanations
Ask for assumptions, sources used, calculation steps that you can double-check, and a brief justification. A request for internal thoughts or long reasoning is not necessary for a useful result. A self-check by the same model does not replace an independent check.
Change and compare
Maintain a fixed set of test cases. Change one instruction and compare the results. This allows you to see if a prompt actually improves or if it only happens to work for the last example. Assign a date or version number to every approved variant.
New practice example
From explanation to a usable result.
Improve one component at a time, so you can see what causes a result to change.
Fictional practice situationFictitious: a AI draft contains all the correct facts, but is 240 words long. You asked for a maximum of 100 words for new volunteers.
View a possible solution
This is how it can be done
Rewrite only the length and phrasing: maximum 100 words, plain language, preserving all confirmed agreements. Do not add anything. Then compare the facts and the word count again.
Why this works
You do not need to replace a usable base entirely. One targeted change is easier to assess than five changes at once.
Make it slightly more difficult
Create a small test overview with the original prompt, the change, and one concrete improvement that you establish.
Extra practice, at your own pace. You do not need to submit anything. This does not count towards an automatic assessment.
Watch and understand · with narration
From vague to useful
With a computer-generated voice, on-screen text and captions in your chosen language. Start and pause whenever you like. Open the video separately.
Read the text version of the video
- A vague request: write an email. The goal, reader and facts are missing.
- Give context: an invitation to a fictional online workshop, on Saturday from ten to eleven.
- Ask for no more than ninety words, plain language and one next step. Leave missing information open.
- Check the time, invented claims and the next step. Then ask for one specific improvement.
Your workplace · learning by doing
Make something. See what changes.
Checking the AI connection… You can already view the written example.
Present your assignment to Ende.
Read the answer, change one thing, and compare.
Previously held
New version
You remain in control
What do you want to improve in your next attempt?
On this device only. AI answers remain only in this page memory. This is self-review, not a grade.
Why this approach works
You do not need to replace a usable base entirely. One targeted change is easier to assess than five changes at once.
A step further
Create a small test overview with the original prompt, the change, and one concrete improvement that you establish.
Try & discover · no obligation
Build a prompt and compare the versions
Switch off context and boundaries. Which instructions disappear from your task?
Vague start
Write a customer email.
Your constructed prompt
Write a friendly customer email about changed opening hours.
Not yet explicitly provided: context and source facts, format and target audience, boundaries and control.
This trial works locally with fictitious data and does not count towards your lesson completion. It starts anew with a new lesson or reload.
A prompt to get you started
Rewrite the text according to these criteria: maximum 100 words, plain language, one concrete next step. Use example A as a style guide: [fictitious example]. Provide the result and then a short check for length, missing information, and added claims. Text: [practice text].
Apply it
Prompt B adds to prompt A: “Ask for missing data.” You want to know if B really works better.
Which comparison yields the most useful information?
- Test A on a difficult question and B on an easy one.
- Test both on the same ordinary question and a case with missing data, according to the same criteria.
- Ask B if it is better than A.
With JavaScript, you can check your approach and receive immediate feedback.
Test your insight
How do you know if an adapted prompt is better?
- It contains more words.
- The model says it is better.
- It performs better on the same test cases according to fixed criteria.
Extra practice
Create two versions of the same prompt. In version B, add one example or control criterion. Describe a normal test and an edge case to compare both.
A asks for a customer response. B adds a fictitious style example and says "ask for missing data". The normal test contains all product details. In the edge case, the delivery time is missing. I check whether B mentions that uncertainty without inventing a date.
