← sources

ChatGPT forgets and ignores rules I give. Is it just ChatGPT, or is this a common trait among all LLMs?

source: reddit (reddit:ChatGPT)published 2026-10-09 09:07ingested 2026-10-10 02:35

retained text

TL;DR: ChatGPT often ignores rules despite promising to follow them. Do other coding AI's do the same?

\---

I have been experimenting with ChatGPT (and Codex, I'm using "ChatGPT" to mean both) for lots of different programming projects, but I am having problems with getting it to obey rules I give it.

\## An example

For example, I set up a "CODING\_RULES.md" document with basic rules for how I organize functions, name variables, commenting style, that sort of stuff. I then tell ChatGPT to always follow these rules, and we commit them to the project repo, and ChatGPT claims to have given itself a rule to always read this document before adding, deleting or modifying code.

It lasts for 3-10 prompts, and then it starts ignoring the prompt in part or whole.

\## One more example

Another example, during architecture discussions, to save on typing, I decide that the abbreviation "DMO" means "document this permanently, push it to the project repo, and then move on to the next question". ChatGPT puts the rule in the document and the thing I forget the name of, the little rule box in the ChatGPT client, and it works ...for several minutes, and then it starts happening: it stops waiting for a prompt instead of continuing down the list of points that we're discussing, and it even claims to document to the repo when it evidently doesn't. It happily admits "no, I didn't do that" when asked, and if I ask "why" or "how" I get nonsense.

I have spent well over 200 hours on just ONE of my AI-supported projects, and at least 3-4 hours of that time has been discussing with ChatGPT what I or it could do differently to stop this from happening. After 40+ years of programming, I have come to expect that if I set a variable to a value, unless something changes that variable, it's going to stay the same (ignoring concepts like scope and EMP corruption of RAM, we're talking on a very fundamental level now). That's simply not true with ChatGPT.

\## Just one more (feel free to skip ahead if you feel you've gotten the gist of it)

The big project I'm working on currently naturally separates into "chapters", and I've put together a step-by-step list of what to do when we finish one chapter and move to the next (do some final documentation, push to repo, tag the commit etc), and ChatGPT will happily either pretend that this list doesn't exist (until I demand "don't trust your memory, look up the actual file we saved it to"), or just skip parts of it.

\## What about the competition?

I suspect this is something that happens with all LLM's, that it's just in the nature of the concept, like the hallucinations and "oh, I decided to delete 90% of your source code, because fuck you, that's why".

Or is it? Those of you that have experienced this specific "rules are *barely* even suggestions" problem AND have used Claude or other AI's intended for coding, does this happen with them too?

document fetched 2026-10-10 00:21 · canonical https://reddit.com/r/ChatGPT/comments/1x1gidf/chatgpt_forgets_and_ignores_rules_i_give_is_it