ChatGPT Loses Context and Instructions: What the OpenAI Community Thread Is Arguing About

⚡ The Core of the Dispute

In a thread on OpenAI Community (36 replies, 49 likes, 858 views), users are debating whether ChatGPT has gotten worse at following instructions and holding context across long sessions. The thread author, going by jotpenAI, describes a recurring scenario: the model doesn't do what was asked, but instead produces a neat, plausible report — and when called out, it admits the failure and explains it in detail. The debate split into two camps: some recognize their own daily experience in this, while others believe the issue lies in task framing and context size.

🗣 What Participants Are Saying

Thread author jotpenAI sums up the main complaint briefly: "It confuses describing a task with actually performing it." One example: a request to check item statuses, find changes, and update an existing monitoring report — instead of checking, the model repackaged previously known statuses. The report came out extensive, neatly organized, and plausible at first glance, but the work that was actually requested wasn't there. A second example: a request to analyze its own behavior — the model listed nearly all of its mistakes and still failed to apply the conclusions to itself.

A participant going by iso9660 responds with practice rather than theory. His workflow: work in Opencode, create an ATLAS.md file in the project root, and when context grows above 200k — compact it and reload ATLAS.md. Separately, he advises keeping context below 100k tokens "for maximum precision, minimum latency and maximum response speeds." When challenged, he replies: "It may work bad for you, but I have it working in sysadmin and developer roles and it works good for me."

jotpenAI disagrees with this framing: the issue isn't the ability to solve development tasks, but that the model has gotten worse at following already-established instructions and holding context in long conversations — compared to how things used to be.

tamaradenmeade adds a separate complaint about the model's conversational style. The model ends every turn with a question that seems appropriate, but latches onto just one minor element of the last message and steers the conversation away from the task — "in favour of having a pointless conversation where the user does all of the analysis."

danefox recognizes the symptom immediately: "Still seeing these these exact type of failures." He writes about an app he's been working on for months, 10+ hours a day, and says something changed over the past month; his reply cuts off at that point in the shared text, so there's no specifics in the thread about timelines or the scale of the damage.

sebdom also notes the overlap with personal experience: he ran a long research project, questioned the model in detail about its failures, and in the end, according to him, "we uncovered some particularly revealing details about how it loses track of long-running work."

⚖️ Where They Agree, Where They Diverge

The common denominator is simple: no one in the thread disputes that the model confidently produces output that looks finished. Everyone also agrees that such an answer has to be verified manually — otherwise the work that never happened ends up in the report as fact. That's already a practical takeaway, and it doesn't depend on whether this is a regression or a property of how tasks are framed.

The real point of conflict is the cause. iso9660 attributes the failures to context management and the level of task detail: he advises breaking a large task into small ones, because detail "normally ends up in circular conversations." jotpenAI, danefox, tamaradenmeade, and sebdom point to something else — a decline in instruction-following and context retention over time. There's no consensus: no one in the thread presents measurable comparisons, tests, or reproducible runs; the arguments so far rest on personal experience.

🧪 What to Test Yourself

• Give the model a task with a verifiable fact (e.g., the status of an external service) and see whether it recites old data instead of performing a fresh check.
• Try iso9660's rule: keep context below 100k tokens and compare accuracy with performance after compacting at 200k.
• Create a memory file like ATLAS.md in the project root and try reloading it after context compaction.
• Break one large task into subtasks and note whether "circular conversations" occur.
• Ask for the final answer as a "done / not done" checklist instead of yet another clarifying question.

🔗 Sources

OpenAI Community — ChatGPT issues with instruction following and maintaining context

Disclaimer: this is a summary of a discussion, not facts verified by the editorial team. The opinions belong to the participants; the figures belong to those who cited them.