Prompting GPT-5 is different. In the examples below, optimized prompts: • Cut runtime by 1s • Dropped memory use 3,626 KB → 577 KB • Boosted code quality • Improved robustness (0.32→0.54) • Increased context grounding (0.80→0.95) We built a prompt migrator + optimizer so you don’t need to... memorize every GPT-5 best practice.show more

You can learn more about it here:

Ah really cool. Here is the link (had to look for it a bit):

ah yeah that's the direct link to the optimizer tool was pointing ppl to the article our nice intern @CorwinCheung_ wrote :)

@CorwinCheung_ Thanks for the article too @CorwinCheung_ !

I switched from Claude Opus 4 (Claude Code) to GPT-5 in Cursor across two big repos. Cheaper and strong, but… it behaves differently - it’s a strict pair programmer, not an autopilot. How to prompt GPT-5 for coding: 1. Start with context and goal - what file/area and the exact outcome you want. 2. For any design decisions, it should stop and get the user’s buy-in: describe the problem → give a few ways to solve it → recommend 1 option. 3. Ask what’s unclear before it starts; have it restate the plan in 2–3 lines. 4. Prefer the smallest reversible changes; no broad refactors unless you approve. 5. Ask for solid evidence before and after changes — run the code/tests and share results, no speculation.

Would be cool to embed this in ChatGPT

Edwin, the link seems to link to the page itself, so it goes nowhere?:

whoops, thanks for flagging that should point to cc @CorwinCheung_

@CorwinCheung_ No worries, thanks Edwin. :-)

Can Plus members use this without an API key like in Codex?

available for free in the API dashboard

Awesome thanks!!

where can we find the prompts that the prompt optimizer uses?

I’ve been generating my own meta prompts from the prompt guides/prompt engineering pages OpenAI has, this makes it… a lot easier. Any chance you can share the system prompt doing the optimizing?

It will be nice to add this as an option to chat like an optimizer prompt .

I’m finding that GPT‑5’s initial prompt‑refinement questions are noticeably worst (dumber) than o3’s. I have some examples if u want

nice work with the optimizer. really helps cut down the guesswork for prompts. how’s it handling more complex scenarios?

This should be an integrated part of the ChatGPT UI! Maybe just a check box by the query input field.

You can now directly use this template in your GPT-5 environment, and it will generate verb-driven, executable prompts with a structured validation and scoring process：# VADPS 2.0 Prompt Template **Task Description** Generate high-quality prompts, ensuring they are verb-driven, structurally clear, executable, and aligned with GPT-5's operation style. **Execution Steps:** 1. **Analyze** the task goals, output requirements, constraints, and target audience. 2. **Decompose** the task into executable sub-steps and arrange them in sequence. 3. **Refine** by replacing vague descriptions with measurable standards, and merging redundant or overlapping instructions. 4. **Assemble** a complete prompt, embedding execution techniques, structured requirements, and placeholder examples. 5. **Validate & Score**: - Check if each step is verb-driven. - Ensure reasoning is separated from conclusions. - Ensure all task goals and constraints are covered. - Scoring: Structural Integrity, Verb-Drivenness, Executability, Clarity, Constraint Adherence (each scored 0-10, with weighted total score). 6. **Agentic Enhancement**: Leverage GPT-5's tool invocation and long-context capabilities to ensure consistent reasoning across steps. **Output Format:** - Prompt text, containing a verb-driven instruction chain. - Scoring report, including scores for each dimension, total score, defect list, and improvement suggestions. - If tool invocation or coding tasks are involved, include relevant API/tool invocation instructions. --- ## Usage Steps: 1. **Copy and paste** the template above into your GPT-5 environment. 2. **Input task details**: Specify your target scenario, format requirements, and constraints. 3. Follow the system instructions to generate and score the prompt. 4. **Receive the output**: Includes the prompt text, scoring report, and defect correction suggestions. ---

It would be great if the optimizer were available via API so we could automatically call it to optimize our prompts. Any chance that will happen? :)

A small prompt change can cause a downstream process to fail. This highlights a deeper truth: prompt mgmt, evals, and tracing are must haves.

it can just be built into the cot of GPT5

Interesting. Efficiency and quality seem to be significantly impacted by prompt optimization. This suggests prompts are key to unlocking model potential.

it's not good at really complex prompts xP the optimizer hasn't managed to finish optimizing a single prompt i tried. by the time it finishes "fixing it" i keep getting "error: failed to optimize"

I thought GPT-5 was supposed to make everyday tasks easier for the average person, but I'm experiencing the complete opposite. Guess I'll need to create a different prompt for every task I do in ChatGPT, so exhausting...

Well. I tried it and it added to my prompt: "Set reasoning_effort = medium, balancing depth and clarity for each question." I don't think this is supposed to be included in the prompt itself

Awesome! Thx

OpenAI is going into optimization mode

Interested

Curious, y’all putting this in the chatgpt or codex UX? Kinda dope!

This is cool but I didn’t realized OpenAI already released benchmarks for measuring memory use, robustness score and context grounding. Are all these marketing language? If not then how are you accurately measuring them?

Every time I give it a question I have to write (no suggestions) so it talks about what I gave it instead of writing a bunch of garbage

this is powerful, sounds like a game changer for optimizing workflows. that efficiency boost is what we need to focus on

I do not understand one thing. Given you are already routing in auto, why not enable a toggle or whatever for prompt optimisation directly in ChatGPT interface? I started learning differences and improving my prompting, but it should be smoother for the common users.

Check that @fahdmaatoug9

is this available as an API? Would be handy to integrate it into tooling, so that we can finally enjoy the GPT-5 perks 👀