When an ai repeatedly misses your preferred result, training can sound like the obvious next step. First, diagnose the failure. A vague request, missing facts, and an unsuitable example can all make a task difficult to evaluate.
Take a fictional enquiry reply. You want the answer to acknowledge the request, list missing details, and avoid promising availability. “Write a professional response” does not express those requirements.
Name the failure
Keep an example of the unwanted output. Mark the exact problem: invented commitment, missing question, wrong format, or unsuitable tone. Different problems may need different changes. Improving the tone will not supply a missing source.
Improve one instruction
Write the job in plain language:
Draft a response to this fictional enquiry. Acknowledge the request, then ask for the missing date and participant count. Do not invent a price or confirm availability. Finish with a request for those details. Keep the answer concise and friendly.
Add the relevant source material separately. Give the assistant a way to report missing information instead of making completion appear compulsory.
Add a useful example
Choose a short example you own or may use. Explain which features matter: a clear acknowledgement, concrete questions, and no unapproved commitment. Remove old offers or identifying information that should not become part of the new output.
An example is not a licence to copy every detail. State which facts belong to the current task and which part illustrates style.
Compare against a fixed check
Use the original task and the revised brief on material you can inspect. Look for the named failure as well as new problems. Keep complete responses and relevant settings. If the improved brief still fails, describe the remaining pattern rather than immediately escalating the whole project.
This exercise may show that the assignment needed better boundaries. It may also show a limitation that requires another approach. Either conclusion is useful when grounded in an actual comparison. Supplying a better prompt or examples is not the same activity as changing a model's learned parameters.
Mika is an editorial pen name used by the KENJTEC editorial team in Hamburg. This article proposes a workflow; it does not report a completed test.
