How to Prompt Claude Sonnet 5 (You've Been Writing Too Much)

I opened my saved prompts folder last week.

340-word instructions. Numbered step lists. "You are an expert." "Think step by step." "Re-read before answering." "Do not hallucinate."

Two years of work. Built for a model that's no longer your default.

Claude Sonnet 5 shipped on June 30. If you're on a Free or Pro plan, you switched to it automatically. No update required. No prompt. It just happened.

And those saved prompts you've been refining for months? They're now telling a model that already thinks on its own exactly how to think. That's not helping it. That's slowing it down.

This guide covers what changed, what to cut, the four-line structure that replaces everything you delete, and where to put the instructions that actually still matter.

Three behaviors are built into Sonnet 5 that you used to have to ask for.

Adaptive thinking is on by default. Every prompt you send now runs with thinking enabled automatically. Sonnet 5 decides how deeply to reason based on your task. A simple lookup gets a simple answer. A complex problem gets real analysis. You don't tell it to think carefully. It already is.

It checks its own output without being asked. Anthropic confirmed this in the launch notes. The Zapier team tested it on a two-part task - update Salesforce account tiers, send a launch announcement - and their senior engineer said: "It finished end to end. That used to stall halfway." Checking its own work is now standard behavior, not something you prompt for.

It plans its own steps. Tasks that Sonnet 4.6 would stop halfway through and hand back, Sonnet 5 completes. Give it a goal. It maps the path. You don't need to write the steps.

When you write a prompt that specifies all of this, you're not adding safety. You're adding interference.

Your old prompts were the right call. They were built for a model that needed the guidance. Sonnet 5 doesn't need the same guidance anymore.

These five things were load-bearing once. They're overhead now.

Sonnet 5 already knows the domain. Role-priming it is like handing a chef a Wikipedia article on cooking before they start. It costs tokens. It adds nothing.

Adaptive thinking is on. Telling it to think step by step is the equivalent of reminding someone to breathe while they're running. The instruction is redundant at best, disruptive at worst. If output feels shallow, the answer isn't this instruction. We'll get to the real fix in a minute.

Sonnet 5 plans better than your numbered list. When you give it explicit step-by-step instructions, you're substituting your planning for the model's. Give it the goal. Let it map the path.

Built in. This is one of the documented behavior changes in Sonnet 5. It checks its own output. You don't request it. You just get it.

This never worked. And Sonnet 5's hallucination rate is measurably lower than its predecessor. The instruction adds nothing except a reminder that you don't fully trust it. Which it ignores.

This is the whole prompt. Not 500 words. Not numbered steps. Four lines.

Line one gives Claude everything it needs to understand context. Line two tells it where to land. Line three hands the formatting and workflow to a saved skill. Line four stops it from guessing when information is missing.

That last line matters more than it looks. Sonnet 5's lower sycophancy means it'll actually ask rather than fill gaps with assumptions. Give it permission to stop and ask, and you'll get better output in fewer exchanges.

I tested this on the research workflow I use before every newsletter edition. I'd been running a 340-word prompt for it for eight months. I ran the 4-line version between two lectures last Tuesday. The output quality matched. The prompt took 45 seconds to write.

Become a paying subscriber to get access to this post and all other premium content.