How to Prompt Claude Opus 5 and GPT-5.6: What Changed in 2026

Claude Opus 5 and GPT-5.6 Work Best When You Stop Overexplaining

Most founders have a folder full of prompts that worked perfectly six months ago. They're still copying them into every new chat, still wondering why the output feels forced or off.

The models changed. The prompting didn't.

Claude Opus 5 launched in July 2026, and Anthropic published new guidance the same day that turned conventional prompting advice upside down. They removed over 80% of the system prompt they'd been using for Claude Code with no loss in quality. GPT-5.6 followed the same pattern: shorter, simpler prompts produce better results than the detailed instruction sets we all learned to write in 2024 and 2025.

This isn't a minor tweak. It's a fundamental shift in how these flagship models reason, and it means the prompts you've been refining for the last two years might actually be making your results worse.

What Changed: Models Got Better at Understanding Intent

The breakthrough isn't that Claude Opus 5 and GPT-5.6 are faster or cheaper. It's that they're better at inferring what you want without you spelling out every edge case and formatting rule.

Earlier models needed guardrails. You had to tell GPT-4 not to make things up, tell Claude 3 Opus exactly how to format a response, and specify tone in painful detail or you'd get something generic. The models were powerful, but they needed tight constraints to stay on track.

Claude Opus 5 and GPT-5.6 can hold your intent across longer conversations and apply judgment to gaps you don't explicitly fill in. That's the core improvement, and it changes everything about how you prompt.

When Anthropic removed 80% of their system prompt, they weren't cutting corners. They were proving that the model had internalized the principles those instructions used to enforce. The model doesn't need to be told "be concise" or "use clear language" anymore. It defaults to that.

Why Your Old Prompts Feel Worse Now

If you've been using the same detailed prompt templates since 2024, you might've noticed the output feeling stiff, overly formal, or weirdly constrained. That's because you're giving the model instructions it no longer needs, and those instructions are now acting like walls instead of guidance.

Here's what happens when you over-prompt a newer model:

• You tell it to "be concise" and it compresses ideas so hard the nuance disappears
• You give it a tone guide and it locks into that tone so rigidly it sounds robotic
• You list formatting rules and it prioritizes structure over substance
• You tell it what not to do and it spends cognitive effort avoiding those things instead of solving the actual problem

The model is doing exactly what you asked. The problem is you're asking for things it would've done better on its own.

This is hardest for people who got good at prompting in 2024. You spent months learning how to write detailed, structured prompts that controlled for every variable. That skill made you effective with GPT-4 and Claude 3. Now it's working against you.

The New Prompting Framework: Start Simple, Add Context Only When Needed

The updated approach flips the old method. Instead of starting with a long prompt and trimming it down, you start with the simplest possible request and only add detail where the model actually needs it.

Here's the structure that works now:

1. State the Task Clearly

Tell the model what you want in one sentence. Be direct. Skip the preamble.

Instead of: "I need you to help me write a blog post. The blog post should be about AI for consultants. It should be practical and actionable. The tone should be professional but approachable. Please make sure to include examples."

Write: "Write a 1500-word blog post on how consultants can use AI to automate client onboarding."

Claude Opus 5 knows what a blog post is. It knows consultants. It knows what "practical" sounds like for that audience. You don't need to define those terms unless your use case is unusual.

2. Add Your Context, Not Generic Instructions

This is where most people still get it wrong. They add rules ("be concise," "use bullet points," "sound professional") when they should be adding context.

Context is the specific information the model can't infer: your business, your audience, your voice, your constraints. Rules are things the model already knows how to do.

Good context looks like this:

• "My clients are fractional CMOs who work with 3-5 companies at a time."
• "Our onboarding currently takes 90 minutes per client and involves three separate forms."
• "I use Kit for email and Airtable for client data."
• "My audience hates jargon and AI hype. They want proof before they'll try anything."

That context tells the model things it couldn't guess. It doesn't tell the model how to write, because the model already knows how to write.

This is the core principle behind Context Training, the method Makeda Boehm teaches at Seed & Society. AI without your context is a brilliant stranger guessing at your business. Once it has your context, it stops guessing and starts working.

3. Let the Model Use Judgment

Older prompting advice told you to close every gap. If you didn't specify a word count, the model might write too much or too little. If you didn't define tone, you'd get something generic.

Claude Opus 5 and GPT-5.6 are better at making those calls themselves. If you ask for a blog post, they'll default to a reasonable length. If you ask for an email to a client, they'll match business-appropriate tone. If you ask for a script, they'll format it correctly.

You still add constraints when they matter. If the email has to be under 150 words because it's going into a text message, say that. If the post needs to be exactly 2,000 words to match your SEO template, specify it. But if the constraint is just "make it good," the model will handle that without you micromanaging it.

Real Examples: Before and After

Here's what the shift looks like in practice.

Example 1: Writing a Client Proposal

Old prompt (2024 style):
"I need you to write a proposal for a consulting engagement. The proposal should be professional and persuasive. It should include a clear scope of work, a timeline, pricing, and next steps. Use bullet points where appropriate. Keep the tone confident but not arrogant. Make sure to address potential objections. The client is a mid-sized B2B SaaS company. They're concerned about implementation time and ROI."

New prompt (2026 style):
"Write a consulting proposal for a 12-week AI implementation project. The client is a 50-person B2B SaaS company. They want to automate their customer onboarding process, which currently takes their team 10 hours per new customer. Their main concerns are implementation time and whether they'll actually see ROI within six months. Pricing is $18,000."

The new version gives the model context: company size, the specific problem, their concerns, your pricing. It doesn't tell the model how to write a proposal, because Claude Opus 5 already knows what a proposal looks like. The output is better because the model has room to structure it in the way that makes sense for this specific situation.

Example 2: Creating a Weekly Email

Old prompt:
"Write an email to my list. The email should be around 300 words. Use a conversational tone. Start with a hook that grabs attention. Include one main teaching point. End with a soft call to action. Don't be salesy. Make sure it's easy to scan. Use short paragraphs."

New prompt:
"Write this week's email to my list. Topic: how to use Claude to draft client proposals in 15 minutes instead of 2 hours. My readers are consultants and fractional executives who are skeptical of AI shortcuts. I want them to try the method, not buy anything."

The difference: the new prompt tells the model who you're writing to, what you're teaching, and what you want them to do. It doesn't dictate structure or tone, because the model can infer those from the context you gave.

When You Still Need Detailed Instructions

Simpler prompts work for most use cases, but there are times when you still need to be specific.

Add detail when:

• You're working in a specialized format the model might not default to (a legal memo, a grant application, a screenplay)
• You need the output to match an existing template exactly
• You're asking the model to follow a process with multiple steps in a specific order
• Your voice or brand has a distinctive style that isn't obvious from context

If you're a speaker who records content and needs to turn it into short-form clips, you'd mention that you use Opus Clip to pull the clips, and that the transcripts need timestamps. That's a specific workflow constraint.

If you're creating an online course and you use AICoursify to structure the modules, you'd tell the model how that platform expects content to be formatted. That's a technical requirement the model can't guess.

The rule is: add instructions when the model needs them to do the job correctly, not when you're nervous about letting go of control.

How to Test If Your Prompt Is Too Long

If you're not sure whether your prompt is overbuilt, try this:

Take your current prompt and cut it in half. Remove every instruction that sounds like general writing advice. Keep only the context and constraints that are specific to your business or this particular task.

Run both versions and compare the output.

Most of the time, the shorter prompt produces better results. The output feels more natural, less formulaic, and more like something a human would actually write. If the shorter version is worse, look at what you removed. The thing that mattered was probably context, not instruction.

Using Claude Opus 5 for Prompting: What Works Now

Claude has always been strong at following detailed instructions, and that hasn't changed. What's different with Opus 5 is that it's also better at working with minimal direction.

If you're using Claude for anything that requires judgment, nuance, or adapting to a situation that's hard to define in advance, lean into the simpler prompting style. Let the model think.

If you're using it for structured output where precision matters, like generating data for a CRM or formatting content for a specific platform, you can still be prescriptive. The model handles both approaches well. The key is knowing which approach fits the task.

One practical pattern: start a conversation with a simple prompt, let Claude respond, and then add constraints as you refine. Instead of front-loading every rule into the first message, you're teaching the model your preferences as you go. That's closer to how you'd work with a person, and it works better with Opus 5 than it did with earlier models.

Using GPT-5.6 for Prompting: What to Adjust

GPT-5.6 follows the same general pattern, but it's slightly more dependent on structure in the initial prompt. If you're asking it to create something with multiple parts (a strategy document, a content plan, a presentation outline), give it the structure up front.

Instead of: "Create a content strategy for my consulting business," write: "Create a 90-day content strategy with weekly themes, three posts per week, and one long-form piece per month. I'm a fractional CFO who works with early-stage startups."

GPT-5.6 is better than earlier versions at staying on topic across long outputs, so you can ask for more in one go without the model drifting or losing coherence halfway through.

Where GPT-5.6 really shines: tasks that require pulling from a wide range of knowledge and synthesizing it into something new. It's better at making connections across domains than previous versions. If you're asking it to apply a concept from one field to a problem in another, you'll get stronger output now than you did a year ago.

What Hasn't Changed: Context Still Matters Most

Better models don't eliminate the need for context. They just handle it more efficiently.

If you've been using AI for a while, you've probably noticed that the first result is rarely the best one. You ask for something, the model gives you a decent draft, and then you spend three or four rounds refining it until it's actually useful.

That's a context problem. The model didn't know enough about your business, your audience, or your goals to get it right the first time.

The fastest way to improve your results with Claude Opus 5 or GPT-5.6 isn't better prompting technique. It's giving the model more context about your world before you ask it to do the work.

This is the difference between using AI as a task tool and building an AI employee that owns a role in your business. An agent completes a task. An AI employee owns a role. The employee has enough context to make decisions, handle exceptions, and improve its work over time without you rewriting the prompt every time something changes.

If you're a founder who publishes content regularly, you don't want to re-prompt your way through every post. You want an AI that already knows your voice, your audience, your positioning, and the outcomes you're trying to create. Once that foundation is in place, the actual prompts get shorter because the context is already loaded.

Tools That Support Better Prompting Workflows

You don't need special tools to prompt Claude Opus 5 or GPT-5.6, but a few platforms make it easier to manage context and refine outputs over time.

If you're creating video or audio content and need to repurpose it into text, ElevenLabs gives you high-quality transcription and voice cloning that you can feed directly into Claude or GPT for editing and reformatting. The cleaner the input, the better the output.

If you're managing content distribution across multiple channels and you need AI to help you adapt the same core message for different platforms, Blotato handles scheduling and formatting so you're not manually reposting everything. You write the content once with the help of your AI, and the tool handles the distribution logistics.

The Bigger Shift: From Prompt Engineering to Context Training

Prompt engineering was the skill that mattered in 2023 and 2024. You had to learn how to structure requests, how to format examples, how to chain prompts together to get complex outputs.

That's still useful, but it's not the bottleneck anymore. The bottleneck now is context. The best prompters in 2026 aren't the ones who write the longest, most detailed instructions. They're the ones who've taught their AI enough about their business that the instructions can be short.

Makeda Boehm, Strategic AI Advisor and Digital Workforce Architect at Seed & Society, frames this as the shift from using AI as a tool to building a digital workforce. Tools require instructions every time. Employees build up knowledge over time. The more context they have, the less you have to manage them.

If you're still prompting from scratch every time you need something, you're working harder than you need to. The goal isn't to get better at writing prompts. The goal is to set up your AI so that the prompts you write can be simpler because the context is already there.

What to Do This Week

If you're still using prompts you wrote in 2024 or early 2025, start here:

Pick the three prompts you use most often. The ones you copy and paste into ChatGPT or Claude every time you need a specific type of output.

For each one, cut everything that sounds like generic writing advice. Remove "be clear," "be concise," "use a professional tone," "make it engaging." Keep only the context: who you are, what you're creating, who it's for, and any constraints that actually matter.

Run the shorter version in Claude Opus 5 or GPT-5.6 and compare it to what you were getting before. If the new output is better, save the shorter prompt. If it's worse, look at what you removed and add back only the piece that the model actually needed.

Do this three times and you'll start to see the pattern. Most of what you thought was necessary instruction is just noise now. The models are better, and your prompting can be simpler.

Frequently Asked Questions

What's the biggest difference between prompting Claude Opus 5 and earlier Claude models?

Claude Opus 5 requires significantly less instruction to produce high-quality output. Earlier models needed detailed guidance on tone, structure, and formatting. Opus 5 infers those elements from context and task type, so your prompts can focus on what's unique to your situation rather than explaining how to write. The model is better at using judgment, which means you can give it more room to adapt instead of locking it into rigid rules.

Do I still need to write long, detailed prompts for complex tasks?

Not usually. Complexity in the task doesn't require complexity in the prompt. Instead of writing longer instructions, focus on giving the model better context about your business, your audience, and the outcome you need. If the task has multiple steps, outline the steps. If it requires a specific format, specify that. But general writing quality, tone, and clarity are things Claude Opus 5 and GPT-5.6 handle well without micromanagement.

How do I know if I'm over-prompting?

If your output feels stiff, formulaic, or like it's trying too hard to follow rules, you're probably over-prompting. Try cutting your prompt in half and running it again. If the shorter version produces more natural, useful output, you were giving the model too many instructions. Another sign: if you're spending more time writing the prompt than you spend editing the output, you're working too hard on the wrong part of the process.

Should I still use examples in my prompts?

Examples are still helpful when you're asking for something the model might not default to, or when your style has specific quirks that are hard to describe. But you don't need as many examples as you used to. One strong example is often enough for Claude Opus 5 or GPT-5.6 to understand the pattern. If you've been including three or four examples in every prompt, try cutting it down to one and see if the quality holds.

Can I still use my old prompts, or do I have to rewrite everything?

You can keep using your old prompts, but you'll likely get better results if you simplify them. Start with the prompts you use most often and test shorter versions. You don't have to rewrite everything at once. As you update your prompts, save the new versions so you're not starting from scratch every time. Over a few weeks, you'll have a cleaner, more effective prompt library that works better with the current models.

What should I focus on if I'm new to prompting these models?

Start simple. State the task clearly, add context about your business and audience, and let the model do the rest. Don't try to control every aspect of the output. Most new users over-prompt because they're nervous about what the model will produce. Trust that Claude Opus 5 and GPT-5.6 are good at their job. Your role is to give them the information they can't figure out on their own, not to teach them how to write.

How is this different from prompt engineering?

Prompt engineering focused on structuring instructions to get consistent, controlled output from models that needed a lot of direction. The new approach is more about context than control. You're teaching the model about your world, not telling it how to do its job. The skill isn't writing better instructions. It's knowing what information the model actually needs and giving it that, without over-explaining the rest.

Does this apply to other AI models, or just Claude and GPT?

The principle applies broadly: newer, more capable models need less prescriptive instruction and more relevant context. Claude Opus 5 and GPT-5.6 are the flagship examples right now, but other models are moving in the same direction. As models improve at reasoning and inference, the prompting style that works best shifts from detailed control to contextual guidance. If you learn to prompt effectively for these models, the approach will transfer to other advanced models as they're released.

Individual results vary. Time savings depend on your business, your tools, and how you manage your AI employees.