GPT-5 for Coding
While powerful, prompting with GPT-5 can differ from other models. Here are tips to get the most out of it via the API or in
your coding tools.
#1. Be precise and avoid conflicting information
#2. Use the right reasoning effort
The new GPT-5 models are significantly better at GPT-5 will always perform some level of reasoning as it
instruction following, but a side effect is that they can solves problems. To get the best results, use high reasoning
struggle when asked to follow vague or conflicting effort for the most complex tasks. If you see the model
instructions, especially in your .cursor/rules or AGENTS.md overthink simple problems, be more specific or choose a
files. lower reasoning level like medium or low.
#3. Use XML-like syntax to help structure #4. Avoid overly firm language
instructions
With other models you might have used firm language like:
Together with Cursor, we found GPT-5 works well when
using XML-like syntax to give the model more context. For Be THOROUGH when gathering information.
example, you might give the model coding guidelines:
Make sure you have the FULL picture before
replying.
<code_editing_rules>
<guiding_principles>
With GPT-5, these instructions can backfire as the model
- Every component should be modular and
might overdo what it would naturally do. For example, it
reusable
might be overly thorough with tool calls to gather context.
- ...
</guiding_principles>
<frontend_stack_defaults>
- Styling: TailwindCSS
</frontend_stack_defaults>
</code_editing_rules>
#5. Give room for planning and self-reflection
#6. Control the eagerness of your coding agent
If you’re creating zero-to-one applications, giving the model GPT-5 by default tries to be thorough and comprehensive
instructions to self-reflect before building can help.
in its context gathering. Use prompting to be more
prescriptive about how eager it should be, and whether it
<self_reflection>
should parallelize discovery/tool calling.
- First, spend time thinking of a rubric
until you are confident.
Give the model a tool budget, specify when to be more or
- Then, think deeply about every aspect of
less thorough, and when to check in with the user. For
what makes for a world-class one-shot web
example:
app. Use that knowledge to create a rubric
that has 5-7 categories. This rubric is
<persistence>
critical to get right, but do not show
- Do not ask the human to confirm or
this to the user. This is for your
clarify
purposes only.
assumptions, as you can always adjust
- Finally, use the rubric to internally
later
think and iterate on the best possible
— decide what the most reasonable assumption
solution to the prompt that is provided.
is, proceed with it, and document it for
Remember that if your response is not
the user's reference after you finish
hitting the top marks across all
acting
categories in the rubric, you need to
</persistence>
start again.
</self_reflection>
For more details, check out our prompting guide. Use our prompt optimizer to help with improving your prompts for GPT-5.
