I would like to report a recurring issue with ChatGPT that affects different kinds of tasks. To reduce errors, I have saved memories about my preferences and decisions, set up custom instructions, established rules for how it should behave, and created numerous guides. I also manually add this prefix before every request:
“Before responding or taking any action, read my entire request and all attachments. Review all my saved memories, custom instructions, established decisions, and relevant conversation context. Apply every memory and instruction that pertains to this request; do not overlook or silently replace them. Identify exactly what I want changed and what must remain unchanged. If an ambiguity could materially change the result, ask a specific question before acting. Do only what I requested. Before delivering, check the result against every requirement and compare it with the original when editing. Correct any failure first. If you cannot meet or verify a requirement, say so clearly and do not claim the task is complete. My request:”
Despite this, ChatGPT continues to treat my requests as permission to make changes I did not ask for, remove important information without consulting me, and sometimes suggest steps or settings that are not available to me. It has also responded without being able to properly inspect some attachments and claimed to have restored or verified content that still contained changes or errors.
What frustrates me most is that, when I point out a specific mistake, it often apologizes and repeats the same pattern: it takes another action I did not request or presents a correction that introduces a new problem. This forces me to check its responses and redo work, despite the steps I have taken to provide clear context and instructions.
Please investigate why it does not reliably follow explicit instructions, saved memories, and continuity guidelines, and improve how it checks results before claiming that a task has been corrected.