As I shared in Monday’s detailed ChatGPT-5 run-through (replay available ​here​; get a free trial membership ​here​ for access), GPT‑5 is OpenAI’s most advanced model yet (no matter how much people yearn for 4o).

One of the stand-out features of ​ChatGPT-5​ is that it’s engineered to be highly steerable.

It can follow very detailed instructions with high accuracy, has fewer hallucinations, and is even better at agentic tool calling. (If all this sounds like a foreign language to you, join our next ​Executive Boot Camp​ to confidently Lead with AI.)

And that’s all great, but a more steerable and capable model also means something else: that high-quality prompts are more important than ever.

In this guide, I’ll share practical techniques for steering GPT‑5’s autonomy and reasoning, maintaining clarity in complex “agent” tasks, and structuring prompts for optimal results.

Why GPT‑5 changes how you prompt

OpenAI built GPT-5 not just to “sound smarter,” but to behave differently under the hood.

I found it enlightening to read through ​developer documentation​ and see how OpenAI asks them to interact with ChatGPT-5. In particular, ​this prompting guide​ on OpenAI’s “Cookbook” developer site gave great insight into the inner workings of ChatGPT.

In short, ChatGPT has moved from a helpful assistant to a capable Chief of Staff. (Sam Altman says it’s a “​PhD in your pocket​”, but honestly, I wouldn’t know what to ask a PhD regardless of their location.)

It, has ‘moved up in the org chart’ as I heard someone say, because of several key improvements and design decisions, namely:

• Depth on demand: GPT-5 can choose to think shallow or deep, like previous ​reasoning models​. A ‘router’ decides which mode to tap, unless you direct it otherwise.

• Verbosity control: A new setting lets you keep answers tight without dumbing them down, if you ask it to.
• Agentic eagerness: It’s primed to perform complex, multi-step processes and call tools like web search, data analysis, and ​custom instructions​ autonomously.
• More context: It handles more input (a.k.a. the “​context window​”) without getting lost, from 32k tokens to 196k tokens.
• Fewer fabrications: ChatGPT-5 ​prevents hallucinations​ by up to 80%, and is better at saying “I don’t know.”

• Rule follower: GPT-5 follows instructions more precisely. Which is great, unless your rules contradict each other.
• Plans and progress: It can tell you its plan and give progress updates along the way, but again, only if you ask it to.

Putting ChatGPT-5’s Design Decisions into Practice

I’m never too focused on the technical background of AI, but understanding these design decisions and knowing how developers are told to handle the model gives us helpful hints into how to get the most out of ChatGPT-5 ourselves.

Below are the executive prompting guidelines that map 1‑to‑1 to what’s new in GPT‑5. To be as practical as possible, I’ve drawn from OpenAI’s official guides and added “field reports” from expert prompters I trust, like Nate Jones and Dan Shipper.

1: Decide fast vs. thorough, in your first sentence

What to say (quick): “Give me a concise 5-bullet answer. If deeper research is needed, don’t do it, but flag the single most critical open question.”

What to say (deep): “Think longer before answering. Explore options and risks, then give me your final recommendation.”

Why it works: You’re signaling the router on generating a quick take versus deep reasoning. Length and depth are separate, so make sure to set both.

Field note: Dan Shipper’s team, who tested GPT-5 hands-on for weeks, ​found​ ChatGPT’s non‑reasoning pass can occasionally hallucinate on tasks that should have been routed deeper and telling GPT‑5 to “think longer” fixed it.

OpenAI’s ​own guidance​ is to use lower effort for quick retrievals and summaries, and dial it up for multi‑step or ambiguous work. (But conspiracy theorists would argue that OpenAI’s interest is token minimization, so take that advice with a grain of salt.)

2: Set eagerness so it neither stalls nor runs ahead

Less eager (draft with detours): “Limit exploration. No web browsing or tools unless essential. Stop after one pass and give your best answer. If uncertain, state the assumption and proceed.”More eager (let it run): “Keep going until the task is fully done. If uncertain, choose a reasonable assumption, proceed, and document it. Only stop when all sub-tasks are complete.”Why it works: GPT-5 is agentic; without an “eagerness” setting, it may over-search or bounce questions back unnecessarily.

Field note: Ethan Mollick ​describes​ GPT-5 as more proactive by default, happy to suggest next steps and even build beyond your brief. Telling it to push forward under uncertainty (with documented assumptions) channels that proactivity productively, while asking it to limit exploration reins it in for speed.

3: Make the rules explicit and consistent

What to say: “Policy order: Compliance & confidentiality > Accuracy > Speed. Safe to do now: draft docs, analyze data I’ve provided, use open web search and cite sources. Ask before: contacting people, signing up for tools, or spending money. Stop & hand back if you hit legal/compliance, need credentials/payment, or facts can’t be verified.”

Why it works: GPT‑5 follows instructions “​with surgical precision​.” Contradictions force it to reconcile conflicts instead of solving the task.

Field note: OpenAI’s prompting guide explicitly ​recommends​ defining “early stop” criteria and giving the model an “escape hatch” (permission to proceed under uncertainty) to avoid over‑searching.

4: Require a mini-plan and a brief progress note

What to say: “Restate my goal, propose a 3-step plan, then execute. For longer tasks, narrate brief progress before delivering the final result.”Why it works: GPT-5 was trained to do better multi-step work when it states a plan and reports progress. This helps you understand what it’s going to do and how it did it.

Field note: In OpenAI’s GPT-5 ​prompting guide​ for developers, the team highlights how tool preambles (plan first, then act; short progress notes) consistently improved long-run reliability in their testing.

5: Separate depth from length (tone is a setting, not a vibe)

What to say: “Tone: candid and direct. Length: 5 bullets unless I say ‘deeper.’ Escalate red flags, but don’t ask permission for low-risk steps.”

Why it works: You can keep answers tight while still opting into deeper reasoning when it matters. (Versus previously, where shorter answers usually meant sacrificing depth.)

6: Feed it context, but without drowning it

What to say: “Here’s the background (short summary). Use it, but don’t quote large chunks. If you’re unsure, say so and list the top 1–2 ways to verify.”

Why it works: Long context is powerful, but reasoning also consumes space. Summaries leave headroom, reduce the likelihood of hallucinations, and avoid confusing the model with overlapping ideas or instructions.

Note: GPT-5 is an excellent multimodal model. Don’t forget that images, screenshots, or any other type of visual material can be used as input sources.

A simple “executive template” prompt for GPT‑5

In our ​Executive Boot Camp​, we start by understanding that AI isn’t a piece of technology. It’s a new colleague. So prompting is really more like delegating.

In good delegation, it’s best practice to highlight your Objective, some Do’s and Don’ts, and the desired Output, what good work looks like.

For delegation to AI, we just add one more layer: the “Character.” This is because AI is trained on all the human knowledge in the world, so telling it which role to take provides crucial context.

Together, these elements form the CO-DO SuperPrompt framework:

But with ChatGPT-5 being more like a highly capable Chief of Staff instead of a junior colleague, more details, like how long to think, how many tools to use, how much to involve you, and how to report back, is a must.

Here’s what an evolved CO-DO SuperPrompt could look like:

• Character: You are my chief of staff. Research our market entry into <industry/geography>. Restate the goal, then propose a 3-step plan before doing anything. For longer tasks, narrate brief progress.
• Objective: If the task is routine retrieval or summary, stay fast. If the topic is ambiguous, multi-step, or high-stakes, think longer before answering.
• Do’s: Proceed under reasonable uncertainty and document assumptions. Use tools when they help, but avoid over-searching. Summarize long inputs so there’s room to think.
• Don’ts: Do not contact people, spend money, or make anything public. If you hit legal/compliance or missing-data limits, stop, summarize, and hand back a decision.
• Output: H2 headings and bullets. Five bullets max unless I say “deeper.”

Everything above aligns with GPT‑5’s trained behaviors: verbosity control, agentic persistence, tool preambles, escape hatches, and conflict‑free rules.

The bottom line: prompting for ChatGPT-5

Prompting GPT‑5 well isn’t just about briefing, it’s about steering:

• Depth ≠ length. Pick how hard it should think and how long it should speak separately. Say both.
• Eagerness is a dial. Give it “escape hatches” and a finish line so it neither over-searches nor hands work back too early.
• Contradictions are silent killers. One policy order; one set of guardrails. Remove conflicts and quality jumps.
• Plan, then progress. A mini-plan and a short progress note make long runs auditable and easier to correct mid-flight.
• Uncertainty is a feature. Ask it to show doubt and offer verification steps; you’ll ship faster and safer.

Those moves map directly to how GPT‑5 is trained and how early users are getting the best results.

What have you tried so far? Let me know, I’ll keep this updated.

Want to see a live walkthrough of GPT-5 and ​ChatGPT Agents​? I'm hosting a free webinar next Tuesday, which will cover that plus 5 powerful mindset shifts to help you go from AI-curious to AI leader. ​Reserve your seat here.