GPT-5’s debut was met with mixed reactions. Some users quickly embraced its capabilities, while others struggled to get the results they expected. Much of that frustration stemmed less from the model’s limits than from the learning curve of adapting to its new prompting dynamics.
To help close that gap, OpenAI released a comprehensive prompting guide for GPT-5. It shows how the model goes beyond incremental improvements, introducing new behaviors that demand fresh prompting strategies. The guide emphasizes agentic use cases, highlighting GPT-5’s strengths in tool calling, precise instruction following, and long-context reasoning.
I’ve pulled out the main takeaways so you don’t have to read the whole thing yourself.
The New Paradigm: Controlling Agentic Eagerness
The most significant shift in GPT-5 lies in what OpenAI calls "agentic eagerness"—the model's propensity to take initiative versus waiting for explicit direction. Unlike previous models that required coaxing to be thorough, GPT-5 defaults to comprehensive exploration, examining every corner of your codebase and checking all available documents.
This thoroughness comes with trade-offs. By default, GPT-5 prioritizes correctness over speed, conducting extensive searches and analysis to ensure accurate responses. But what if you need faster, more targeted results?
The solution lies in adjusting the "reasoning effort" parameter, available in the Playground, API, and through model selection in ChatGPT. Lower reasoning effort reduces exploration depth while improving efficiency and latency. Many workflows can achieve consistent results at medium or even low reasoning effort, translating to fewer tokens, reduced tool calls, and faster, cheaper responses.
Fine-Tuning Agentic Behavior
OpenAI's guide provides specific templates for controlling how GPT-5 explores problems. For rapid context gathering, the recommended approach includes:
Context Gathering Framework:
Goal: Get enough context fast, parallelize discovery, and stop as soon as you can act
Methods: Start broad, then fan out to focused subqueries in parallel
Early stop criteria: Stop if you can name exact content to change or if top hits converge 70% on one area
For developers building agentic systems, this granular control extends to setting tool call budgets. You can instruct the model to use "an absolute maximum of two tool calls" for rapid responses, or remove all limits for comprehensive analysis.
The Persistence Spectrum
On the opposite end lies "persistence mode," where GPT-5 continues working until problems are completely resolved. The guide provides this template for maximum thoroughness:
"You are an agent. Please keep going until the user's query is completely resolved before ending your turn. Never stop or hand back to the user when you encounter uncertainty. Research or deduce the most reasonable approach and continue."
This approach tells GPT-5 to make reasonable assumptions and document them rather than asking for clarification, enabling autonomous problem-solving for complex tasks.
Tool Preambles: Transparency in Action
One of GPT-5's standout features is its ability to narrate its actions through "tool preambles"—real-time updates about what it's doing and why. This transparency prevents the model from disappearing into lengthy tasks without communication.
The system can be configured for various update frequencies, from detailed explanations of every tool call to brief upfront plans. Here's how it works in practice:
Reasoning: Determining weather response. I need to answer the user's question about weather in San Francisco. I'm going to check a live weather service to get current conditions.
[Tool Call: Weather API]
This transparency proves crucial for maintaining user confidence during extended agentic workflows.
API Evolution: The Responses Endpoint Advantage
OpenAI recommends using the newer Responses API over the traditional Chat Completions endpoint, citing statistically significant improvements. In benchmark testing, the Responses API scored 78.2 compared to 73.9 for Chat Completions on Tau-Bench.
The performance gain stems from context reuse across API calls, enabling improved agentic flows, lower costs, and more efficient token usage. The model can refer to previous reasoning traces, conserving chain-of-thought tokens and eliminating the need to reconstruct plans after each tool call.
Optimizing for Code: Lessons from Cursor
As an early alpha tester, the Cursor team discovered several important nuances when integrating GPT-5 into their coding environment. Their findings reveal the model's unique characteristics and how to work with them effectively.
Initially, Cursor found GPT-5 produced verbose outputs with frequent status updates that disrupted user flow, despite generating high-quality code. The solution involved setting the verbosity parameter to "low" for text outputs while encouraging verbose coding tool outputs specifically.
The Frontend Advantage
GPT-5 shows particular strength in frontend development, especially with specific technology stacks. OpenAI recommends these frameworks for optimal results:
Recommended Stack:
Framework: Next.js with TypeScript and React
Styling: Tailwind CSS, shadcn/ui, Radix themes
Icons: Material Symbols, Hero Icons, Lucide
Animation: Framer Motion
Typography: Sans serif fonts including Inter, Geist, Manrope
The guide emphasizes choosing popular languages and frameworks—advice that extends beyond GPT-5 to AI coding in general.
The Self-Reflection Revolution
For one-shot web application development, OpenAI recommends an innovative approach: having GPT-5 create its own excellence rubric. This self-reflection technique leverages the model's planning and evaluation capabilities:
"First, spend time thinking of a rubric until you are confident. Think deeply about every aspect of what makes for a world-class one-shot web app. Create a rubric with five to seven categories. Finally, use the rubric to internally iterate on the best possible solution."
This approach often yields stronger results by tapping GPT-5’s ability to self-evaluate and refine its output.
Managing Model Intensity
Interestingly, Cursor discovered that techniques used to encourage thoroughness in previous GPT models had counterproductive effects with GPT-5. The new model's natural inquisitiveness meant that prompts designed to maximize exploration actually created excessive tool usage.
The solution involved softening language around thoroughness and removing "maximize" prefixes, allowing GPT-5 to make better decisions about when to rely on internal knowledge versus external tools.
Parameter Fine-Tuning
Beyond reasoning effort, GPT-5 offers several other adjustable parameters:
Verbosity Control: Influences the length of final answers (not thinking). This parameter is separate from reasoning effort and affects only what users see.
Minimal Reasoning: The fastest option that still benefits from the reasoning model paradigm. Best for latency-sensitive applications where extended thinking isn't worth the time cost.
Instruction Following: GPT-5 follows prompts with "surgical precision," which can cause issues with contradictory or undefined instructions. The guide recommends having AI help write and review prompts to identify conflicts.
The Optimization Tool Revolution
OpenAI's new prompt optimization tool represents a significant advancement in AI interaction. Available in the Playground, it analyzes developer messages and suggests specific improvements with detailed reasoning.
The tool explains why each change enhances performance, teaching users effective prompting techniques. For example, it might suggest adding upfront checklists for clearer planning or specifying when to use standard libraries versus external packages.
Users can also request specific modifications, such as asking GPT-5 to "explain everything in detail as it builds," and see exactly how the optimization tool implements these changes.
Practical Implementation
The guide's recommendations reflect a broader shift in AI interaction—from coaxing models to perform to managing their natural capabilities. GPT-5's default thoroughness means the challenge often lies in directing rather than encouraging its efforts.
For developers building with GPT-5, this translates to several key strategies:
Define clear stop conditions and safe behaviors
Use tool call budgets for controlled exploration
Leverage the Responses API for improved performance
Employ self-reflection techniques for complex tasks
Choose recommended technology stacks for optimal results
Looking Forward
GPT-5's prompting guide reveals a model designed for agentic workflows—autonomous systems that can plan, execute, and iterate with minimal human intervention. The techniques outlined here represent early best practices for a new generation of AI interaction.
As organizations integrate GPT-5 into their workflows, understanding these prompting strategies becomes crucial for realizing the model's potential. The difference between struggling with GPT-5 and mastering it often comes down to recognizing that this isn't just a more powerful version of previous models, it's a different tool that requires new approaches.
The mixed reviews that greeted GPT-5's launch likely reflect this learning curve. Those who approached it with fresh strategies found a powerful, capable model. Those who applied old techniques to new capabilities found frustration. OpenAI's prompting guide provides the roadmap for bridging that gap, transforming GPT-5 from a confusing upgrade into the agentic AI system it was designed to be.
Here’s a link to the prompt guide: https://cookbook.openai.com/examples/gpt-5/gpt-5_prompting_guide
