The prompt said ten panels. It sent 2, or 55.
The Auto-generate buttons for a dashboard's title and description tell the model it is shown ten panels. The selection kept the wrong end of each list: a 12-panel dashboard sent 2 panels, a 60-panel one sent 55. A missing backslash also joined panels with a literal n instead of a new line.
We ran the same selection on 1,000 dashboards from the grafana.com library. About 55% gave the model the wrong set of panels: half too few, half too many. On the largest, one title used about 47,800 input tokens instead of about 600.
This is a correctness fix, not a cost story: the token side is worth $0.10–0.56 per 1,000 clicks at GPT-4.1's list price. The model now sees what the prompt says it sees.
- withDescriptions.slice(10)+ withDescriptions.slice(0, 10) …- .slice(10 - withDescriptions.length)+ .slice(0, 10 - withDescriptions.length) …- .join('n');+ .join('\n');The same thing in your product
Not every finding is about cost. When we read the code that builds a prompt, we also check that the model is sent what the prompt says it is sent. Here that was three lines.