Reported by @candid dirge
Created an assistant in the Playground to generate YouTube scripts from 4 text inputs. Added structured system instructions with bullet points and conditional logic. Supplied inputs in plain English format (e.g., "Agency Name: X"). Assistant returned blank output. I refreshed the thread, saw brief success, then blank output resumed. I modified inputs and simplified instructions—only a single-sentence instruction worked reliably.
The assistant should parse the structured instructions and return a full Q&A-formatted YouTube script based on the given inputs, even when multi-step instructions or formatting are present. It should execute consistently across runs without requiring thread refreshes or instruction simplification.
When structured instructions are used, the assistant often returns a blank response. After refreshing the thread, output appears briefly but fails again on the next run. Removing formatting or reducing the instructions to a one-liner causes it to work, suggesting the assistant fails silently when parsing detailed instructions.
Windows 11, Chrome 115, ChatGPT Playground (Assistants), GPT-4.1, Latest API Version