We spent a morning recently sitting with the founder of a pretty technical company, working through workflows and prompts that both sides had been building over the past few weeks. His prompts were more technical. Ours were structured more along the lines you would expect from someone who has spent a lot of time refining prompts for consistency.
The whole point of the session was to combine those two approaches and get to the simplest version that would perform the same way every time. He was sold on Anthropic because of their transparency, and he felt like they were the better company between the two major models. Web Experts tends to use a dozen different tools and looks at it a little differently, because some are better for certain tasks and others fall short depending on what you are trying to do.
The Task Was Simple
The task itself was straightforward: analyze a PDF and generate a clean Excel or CSV output. The prompt was detailed and clearly defined. It outlined the structure, the rules, and exactly what the output should look like.
And still, the results were inconsistent. Same prompt, same document, different outputs. Sometimes it followed everything perfectly, and other times it would skip fields, change formatting, or drift away from the instructions entirely.
Where Team Workflows Broke
What really caught us off guard was when we tried to turn this into a team workflow. We created a project, put a strong prompt behind it, and assumed all we had to do was attach the PDF to get it started. We had seen it work like that before.
But then we would attach the file, and it would ask what we wanted to do, like it had no context at all. The instructions were there and the file was there, and it still did not connect the two. That kind of inconsistency is hard to work around when you are trying to build something that other people are going to rely on.
Splitting The Prompt
Even when we did get it to run the way we expected, it still was not reliable. The same setup would produce different results from one run to the next, which is the exact opposite of what you want in a workflow other people depend on.
At one point we suggested breaking the prompt into two parts, one focused on gathering and structuring the data and the other focused on building the report. That helped a little from a clarity standpoint, but the outputs were still inconsistent. At that point it became clear the issue was not just the prompt; it was the execution.
Reliability Over Capability
At a certain point it becomes less about capability and more about reliability. When you ask for something specific and get inconsistent results, there really are not a lot of options left to try.
After seeing that play out a few times, we made the decision to move the workflow over to OpenAI's platform. It simply produced more consistent results for this type of structured task. This is not about one being better than the other across the board, but when you are building something that needs to work the same way every time, consistency wins.