Improve a step and check the effect

Keep a known starting point.

Save the old step and a sample input. Say what should improve and what must stay the same.

Ask for a focused change, then try it with the same facts. Compare both results with your goal and source.

One better result is useful evidence, not proof that the problem is gone. Try another relevant case, such as a longer list of required facts.

If another known error affects the task, fix it too. Keeping everything else unchanged is not a reason to keep a broken step.

If the routine still fails, keep using a simpler method you can check. You do not have to keep asking AI.

Revise the draft step, keep the check.

Made-up routine: collect the current garden facts; ask for a full warm note; check it against the plan; choose whether to send it yourself. The last reply had four paragraphs. A short note is wanted, with all facts kept.

Repair a packing-list step.

Sample source: hat, water, lunch, towel, keys, bus pass, phone. The old step asks for a five-line packing list. Its reply drops keys and the bus pass. Write a new step and a check.