It’s packed with real techniques used by Cursor, OpenAI teams, and devs building agentic tools.
Here are the top 10 techniques you can steal today 🧵
1/ Be precise and avoid conflicts
- GPT-5 is excellent at following instructions
- But vague or contradictory inputs slow it down
- Clear, specific directions = better results
2/ Match reasoning effort to the task
- Use higher reasoning for complex, multi-step problems
- Use lower reasoning for simple or fast tasks
- Efficiency comes from aligning effort with complexity
3/ Control model eagerness
- GPT-5 can either act fast or explore deeply
- Reduce eagerness when you need speed and fewer tool calls
- Increase eagerness when you want autonomy and persistence
4/ Use tool preambles
- Ask GPT-5 to share a quick plan before acting
- Improves transparency in long tasks
- Easier for users to follow what’s happening
5/ Reuse reasoning context
- The Responses API lets GPT-5 carry over past reasoning
- Cuts costs and latency by avoiding repeated work
- Keeps multi-step workflows consistent
6/ Control verbosity
- Reasoning effort = how deep GPT-5 thinks
- Verbosity = how much detail it gives in the final answer
- Adjust both to fit the task
7/ Build with default stacks GPT-5 excels at
- Next.js + React + TypeScript for frontend
- Tailwind, shadcn/ui, Radix for styling
- Motion + Lucide/Material for polish
8/ Enforce house rules in codebases
- Define clarity, reuse, and consistency standards
- Specify your stack, directory structure, and design system
- GPT-5 then adapts code to match your repo naturally
9/ Use minimal reasoning for speed
- New fastest mode in GPT-5
- Great for latency-sensitive tasks
- Works best if you add planning or summaries to keep it on track
10/ Metaprompting
- GPT-5 can critique your instructions
- Ask it what to change to get closer to your goal
- Iterating like this improves reliability fast
