Six tips for coding with GPT-5:

First tip use Claude.

Full PDF:

you're absolutely correct! great tips for getting better results from ai coding.

tip 1: don't

Please refrain from posting videos and photos. I cannot translate into Japanese, so I cannot understand the content of your posts😭

Can we make it faster? Latency and TPS? At this point this is so much slower than GPT-4.1

is there a way to have you not inject a formatting prompt over the API? it's conflicting with my instructions and causing problems:

damn read it as "sex tips for coding with gpt-5" and thought it's all over

Tip 1. Use Claude

You guys have lobotomized GPT-5. The current model is nothing like the GPT-5 from its release day. Now it’s faster and cheaper just days after launch? Bringing back old models? Is that just a coincidence? I doubt it. Quantized GPT-5 is the answer. Why make it faster and cheaper only days after release why not before? Something fishy is going on.

GPT5 works better with TAL than with XML.

Why don’t you apply it to codex prompt?

Can you share examples with GPT-5 for the OpenAI agents sdk please. Curious about the reasoning effort in Agentic frameworks

Tip 0 : Use Claude sonnet instead

@hive_echo @grok summarize please

@fscozano look it up man

🧠 GPT-5 is sharper—but less forgiving. If your .cursor/rules or files contain overlapping, vague, or nested instructions, expect silent chaos. We’ve seen agent behaviors break due to unresolved directive conflicts. 🛠️ Our fix? 1. Flatten rule hierarchy 2. Separate global vs agent-level logic 3. Run instruction validation passes before deployment GPT-5 rewards clarity. Build your systems like you’re writing for a very smart—but very literal—mind. #AI #GPT5 #AgentOps #PromptEngineering

Is this built into the codex cli system prompt, or just something to add to

Soon we will be writing in code so the dumb LLM understands us

Tldr: Our new model sucks so use our old ones.

I can’t even get it to code simple power-shell monitoring script without it getting confused in simple syntax errors

this simply ain’t it guys this isn’t the 90s

Your thread is everybody's favorite! #TopUnroll 🙏🏼@CJVegas9 for 🥇unroll

Please return gradient in iOS app

The thing model’s quality depends upon is iterative feedback. Big companies like OpenAI have been hiring scientists and people from different domains just to train the model. These tips are sooo useful and, like with all the “trainers”, it’ll take time to make it better

Nobody uses GPT 5 to write code.. Every engineer I know uses Claude 4.

Ok then, we need per model system prompts in the web UI then. Without this I can't update my prompt with these things in mind without wreaking it for the 4o models. I'm guessing this is why you originally removed the 4 series.

Looks like we have to learn prompt engineering again to use gpt 5 specially.

GPT-5 as coding partner, not coder replacement - the best prompts teach the model your thinking.

coding with gpt5

I've found pair programming with GPT-5 to be incredibly efficient - it catches edge cases I'd miss after hours of debugging.

Thank you, that was helpful - it really seems that new models new new prompting guidelines

those formatting prompts can be a pain. they really mess with custom instructions. got to find a workaround.

As much as I'd like to agree I fed the same code request into ChatGPT and Grok. Grok worked on first run. ChatGPT had errors. Pretty typical for my experience with ChatGPT.

I tested it with code against gemini 2.5 Pro with a task.