Shorter does not always mean the same thing
You ask for a shorter product description and get a slogan. A comparison of two solutions turns into two summaries. In situations like these, first describe the finished result: one paragraph, a table of differences or a step-by-step plan. It is a small change, but at least we have something to compare the answer against.
In our workshop, you choose a task, audience and format. The tool assembles a prompt template from those settings in the browser. It does not connect to a model. You can edit the generated text by hand; do that last, since changing the fields above rebuilds the template and replaces your edits.
Without material, there is nothing to shorten
Replace the brackets with the actual content. Shortening needs an original. Comparing needs descriptions and the criteria you will judge them by. A plan needs a goal, deadline and available time. Choosing a format alone does not tell the model what to fill it with.
For the first trial, we prepared a description of a fictional service. It combines orders from three stores into one list, lets an employee change an order’s status and produces a report once a day. We want all that in 16 words without losing any information. Paste the instruction and text below into the model you use.
Sixteen words. Where did the report go?
“Copy prompt” puts the instruction on the clipboard. If the browser blocks copying, the workshop selects the text; use Ctrl+C or ⌘C. Keep the model’s answer beside the original rather than comparing from memory.
Count the words and check every detail. Are there three stores? One order list? Can the employee still change the status, and is the report produced once a day? A short answer can lose that last part. It might also add “real-time notifications.” They sound good, but we never gave our fictional service that feature. The description needs correction even if it fits the limit.
Change one thing at a time
If the reporting frequency disappeared, point the model to that omission. If the description sounds too formal, clarify who will read it. Keep the same material and change one condition in the prompt. It becomes easier to see whether the correction worked.
Before treating the prompt as a finished template, give it two more texts: one with more data and another with missing information. Watch what happens to the numbers and whether the model admits what it does not know. “Do not invent facts” is an instruction, not a guarantee of a correct answer.
What should you keep for next time?
Copy the instruction and an example of a good answer into a note. You can take the workshop itself through “Download HTML” and open it offline. That copy has the initial settings; your current manual edits will not be included.
To practice extending it with AI, add a “prepare interview questions” task. Ask to preserve local operation and check that the new option changes the prompt. Then fill it with the conversation you actually plan to have. It is easier to notice a missing topic, goal or detail about the interviewee than when looking at an empty template.
Sources and further reading
The sources behind this article. Instructions and example scenarios are suggestions from our editorial team.