How to Fix a Custom GPT That Ignores Its Instructions

A common frustration: you wrote clear instructions, but your custom GPT still gives long answers, ignores the format, or forgets a rule halfway through. Most of the time the cause is fixable in the configuration. This guide is a checklist for finding and fixing the drift.

What you need

• The custom GPT that is misbehaving, in edit mode
• A specific example of the wrong behavior to test against
• Ten minutes to iterate

Step 1: Reproduce the failure in Preview

Open the GPT in the editor and use the Preview pane to trigger the bad behavior on purpose. You need a repeatable failing case so you can tell whether each fix actually helped.

Step 2: Look for conflicting or buried rules

Re-read the instructions for contradictions, like asking for both thorough and very brief. Long instructions also bury important rules. Move the must-follow rules to the top and remove anything that competes with them.

You are a helpful assistant that explains things thoroughly and
in depth with lots of examples and context for the reader, and
you should also try to be concise when possible.

Default to short answers: 1 to 3 sentences.
Only expand when the user explicitly asks for detail or examples.
If the user says "one sentence", reply with exactly one sentence.

Step 3: Make rules concrete and testable

Models follow specific, checkable rules better than vague ones. Replace be concise with a number, and replace nice tone with the exact behavior you want. Negative-only rules (never do X) work better when paired with what to do instead.

Step 4: Check Capabilities and Knowledge for interference

Sometimes the drift comes from elsewhere. A web-search or code capability the GPT does not need can pull it off task, and a large knowledge file can crowd out your instructions. Turn off unused capabilities and trim oversized files, then retest.

Result

After tightening priorities, making rules concrete, and removing interference, your GPT follows its format consistently, and you have a repeatable test to catch regressions the next time you edit it.