Most people think prompting is just typing smart things into a chat box.
But OpenAI just released its official GPT-5 Prompting Guide - and it’s 100x deeper than that.
Here’s what they recommend (and why it matters): 👇
But OpenAI just released its official GPT-5 Prompting Guide - and it’s 100x deeper than that.
Here’s what they recommend (and why it matters): 👇
1/ GPT-5 is not just an upgrade.
It’s a massive leap in:
• Tool calling
• Instruction-following
• Agentic workflows
• Coding intelligence
But to get peak performance? You have to prompt it the right way.
It’s a massive leap in:
• Tool calling
• Instruction-following
• Agentic workflows
• Coding intelligence
But to get peak performance? You have to prompt it the right way.
2/ Want GPT-5 to think less and move faster?
Set a lower reasoning_effort.
Use tight prompts like:
“Give the best answer as fast as possible. Use max 2 tool calls. Proceed even if not 100% sure.”
This reduces overthinking, latency, and tool spam.
Set a lower reasoning_effort.
Use tight prompts like:
“Give the best answer as fast as possible. Use max 2 tool calls. Proceed even if not 100% sure.”
This reduces overthinking, latency, and tool spam.
3/ Want it to think more and solve big tasks on its own?
Use this agentic persistence prompt:
“Keep going until the problem is solved. Don’t hand back to the user. Research and act if unsure.”
It shifts GPT-5 from "assistant" to "autonomous agent."
Use this agentic persistence prompt:
“Keep going until the problem is solved. Don’t hand back to the user. Research and act if unsure.”
It shifts GPT-5 from "assistant" to "autonomous agent."
4/ For complex workflows, use Tool Preambles.
Tell GPT-5 to:
• Rephrase your goal
• Outline a step-by-step plan
• Narrate actions as it executes
Result: You always know what it's doing - and why.
Tell GPT-5 to:
• Rephrase your goal
• Outline a step-by-step plan
• Narrate actions as it executes
Result: You always know what it's doing - and why.
5/ Use the Responses API for long agentic flows.
It remembers reasoning between tool calls.
In benchmarks:
• Tau-Retail score jumped from 73.9% → 78.2%
• Less token waste
• Faster, smarter responses
This is essential for apps with multiple steps.
It remembers reasoning between tool calls.
In benchmarks:
• Tau-Retail score jumped from 73.9% → 78.2%
• Less token waste
• Faster, smarter responses
This is essential for apps with multiple steps.
6/ GPT-5 is also a coding machine.
It can:
• Build fullstack apps from scratch
• Refactor huge codebases
• Follow your team’s code style
Want better results? Use prompts with:
• Project structure
• Design principles
• Naming conventions
It can:
• Build fullstack apps from scratch
• Refactor huge codebases
• Follow your team’s code style
Want better results? Use prompts with:
• Project structure
• Design principles
• Naming conventions
7/ Cursor’s trick: Dual verbosity.
• In normal chat? Keep it short.
• In code tools? Encourage verbose, readable outputs.
They tuned prompts to say:
“Use clear names. Comment where helpful. Avoid one-liners unless asked.”
Clarity > cleverness.
• In normal chat? Keep it short.
• In code tools? Encourage verbose, readable outputs.
They tuned prompts to say:
“Use clear names. Comment where helpful. Avoid one-liners unless asked.”
Clarity > cleverness.
8/ Don’t over-prompt GPT-5.
Cursor found that prompts like “Be thorough” made GPT-5 overuse tools - even when it already had the answer.
Instead, guide it softly:
“If you're unsure, check. But trust internal knowledge first.”
Less is often more.
Cursor found that prompts like “Be thorough” made GPT-5 overuse tools - even when it already had the answer.
Instead, guide it softly:
“If you're unsure, check. But trust internal knowledge first.”
Less is often more.
9/ Final tip: Let GPT-5 prompt itself.
Want better prompts?
Ask it:
“What’s missing in this prompt that caused [undesired behavior]? How would you change it?”
It's scary good at optimizing its own instructions.
Want better prompts?
Ask it:
“What’s missing in this prompt that caused [undesired behavior]? How would you change it?”
It's scary good at optimizing its own instructions.
10/ Key takeaway:
Prompting GPT-5 isn’t about tricking the model.
It’s about steering a super-intelligent agent with clear goals, constraints, and structure.
The better you prompt, the more it feels like magic.
Prompting GPT-5 isn’t about tricking the model.
It’s about steering a super-intelligent agent with clear goals, constraints, and structure.
The better you prompt, the more it feels like magic.
10x your prompting skills with my prompt engineering guide
→ Mini-course
→ Free resources
→ Tips & tricks
Grab it while it's free ↓
godofprompt.ai/prompt-enginee…
→ Mini-course
→ Free resources
→ Tips & tricks
Grab it while it's free ↓
godofprompt.ai/prompt-enginee…
Save this thread. It’ll save you hours.
1/ Follow me @godofprompt for more AI tips.
2/ RT & Bookmark this post:
1/ Follow me @godofprompt for more AI tips.
2/ RT & Bookmark this post:
https://x.com/godofprompt/status/1959298059428397271
• • •
Missing some Tweet in this thread? You can try to
force a refresh
