fleet thorn
ChatGPT should recognize and enforce known character limits when generating text for its own instruction/configuration fields.
This applies to Project Instructions, Custom Instructions, Custom GPT instructions, agent/system setup fields, and any other ChatGPT field with a known hard limit.
Observed behavior:
- ChatGPT can generate “paste-ready” replacement instructions that exceed the field limit.
- This can happen even when the user states the exact limit.
- The model may still present the answer as ready to paste.
- The UI then rejects the text.
Example:
The ChatGPT Projects Instructions field has a known 8,000-character hard limit. ChatGPT can still generate over-limit Project Instructions, causing the UI to reject the text with: “Project instructions cannot be longer than 8000 characters.”
The same issue can happen with other instruction fields, including Custom Instructions.
The trust issue is direct: if ChatGPT knows the target field, the hard limit, and that the user intends to paste the output there, it should not produce text that its own UI rejects.
For serious workflows, this creates a confidence problem: how can ChatGPT reliably help configure ChatGPT if it does not recognize the limits of the fields it is configuring?
Suggested improvements:
- Enforce known limits before presenting output as paste-ready.
- Count spaces and line breaks.
- Show the exact final character count.
- Warn when requested content cannot fit.
- Suggest moving long policy/workflow/reference material into project files or scoped knowledge sources.
- Add live character counters to instruction fields.
- Clarify what belongs in short instruction fields versus long-form reference files.
Environment:
ChatGPT Projects; Project/Custom/GPT instruction fields; ChatGPT Pro; GPT-5.5 Thinking / Heavy / Pro; macOS; Atlas and chatgpt.com.
stable basin