I ran the same planning prompt over 10 app ideas and the last step was hiding real work in 8 of them

I have a prompt I use to turn an app idea into an ordered build plan. It's been working well enough that I stopped reading the output carefully, which is how this went unnoticed for a while.

Last week I ran ten different web app ideas through it and actually read what landed in every single step. Not the summary, the contents.

In 8 of the 10 plans, the final step was named something like "polish" or "final touches" and contained real work. A request counter. Copy to clipboard. Order status tracking. Reports for a team. In two of them it held an entire dashboard that appeared nowhere else in the plan, so if you followed the plan top to bottom you'd hit the last step and discover you still had a dashboard to build.

The reason is the cap. My prompt asks for a bounded number of steps. Work that doesn't fit inside that number doesn't get dropped, and the model doesn't tell you it dropped anything either. It gets folded into the last step under a name that sounds like it means nothing. Every individual step reads fine. The plan reads fine. The problem only shows up when you count what's actually inside it.

Since then I've been checking two things on any plan a model gives me: whether the last item is heavier than the ones before it, and whether anything I mentioned in the original description is missing from every step title. Both take about a minute and both have caught things.

Curious whether people who use step counts in their prompts see the same thing, or whether this is specific to how I worded mine. If you ask for a plan without a number at all, does it come back more honest or just longer?