I have a confession. For a long time, nearly every prompt I sent to Claude ended with some version of “think step by step.” It felt like a magic spell. If the answer came back weak, I’d tack on “think harder” and hit enter again.
So when I found a post from an AI professional saying this habit is now outdated, I stopped scrolling. The author opens with a line that sums up the last year: “Claude is FINALLY back.” ChatGPT beats Claude. Then Claude beats ChatGPT. According to the post, that back-and-forth is speeding up.
The creator’s main point is simple. The better you prompt, the better your results. The catch is that prompting the new Claude, Opus 5.5, is drastically different from prompting the old one. Some tricks many of us picked up are now dead weight.
Why your old prompts might be slowing you down
Older models often needed a push. If you didn’t tell them to reason through a problem, they’d sometimes jump straight to a shallow answer. That’s why phrases like “Think step by step,” “Think hard,” and “Maximum reasoning” spread everywhere. The author even admits to sharing one of those tricks back in the day.
The post says that’s no longer how Claude works. Claude always thinks now, and it decides for itself how much. A simple question gets a quick answer. A tricky one gets deeper reasoning. You don’t have to beg for it.
The detail that surprised me most comes from Anthropic’s own testing, which the expert highlights:
In Anthropic’s tests, deleting “think carefully” lines made Claude start answering sooner, with no clear drop in quality.
So the answers come faster and the quality holds. Those extra lines were basically a tax on every prompt you sent.
The first rule: delete “think harder”
This is Step 1 in the guide the author put together after spending hours on how to prompt Opus 5.5, covering best practices and the mistakes to avoid. It’s also the easiest win on the list. Here’s how I’d put it to work, one move at a time.
- Audit your saved prompts. Go through your prompt library, custom instructions, Projects, and any system prompts you use at work. Search for “step by step,” “think hard,” “think carefully,” and “maximum reasoning.” Why: these phrases hide in templates you wrote months ago and forgot about.
- Delete the thinking instructions. Cut them completely. Don’t swap in a softer version. Why: based on Anthropic’s tests, Claude starts answering sooner without them and the quality stays about the same.
- Spend those words on context instead. This is my own takeaway. Tell Claude who the output is for, what the goal is, and what a good result looks like. Why: Claude already handles the reasoning. What it can’t guess is your situation.
- Run a quick side-by-side test. Take one real task and run the old prompt and the cleaned-up one. Compare speed and output. Why: it’s worth checking the claim on your own work before you rewrite everything.
- Update your templates. Once the lean version holds up, save it as your new default. Why: one cleanup pays off on every prompt you send after it.
What this looks like in practice
A before and after makes the idea click. Here’s a typical old-style prompt:
Before: “Think step by step and use maximum reasoning. Write a cold email to a marketing agency owner about our analytics tool.”
Notice that half of it is about how to think and almost none of it is about the actual job. Here’s the leaner version:
After: “Write a cold email to a marketing agency owner with a 10-person team. They spend hours every month building client reports by hand. Our analytics tool automates those reports. Keep it under 120 words, friendly tone, with one clear ask for a 15-minute call.”
The second prompt has zero thinking instructions and much more useful information. From what the author describes, that’s exactly the trade Opus 5.5 rewards.
Why this shift matters
I think this tip points to a bigger change in how we work with AI. Reasoning is now built into the model instead of being something we switch on with magic words. So prompting is moving away from coaxing and toward briefing.
It’s a bit like the difference between managing an intern and working with a senior colleague. With the intern, you spell out how to think through the problem. With the senior person, you explain the goal and the constraints, then get out of the way. Opus 5.5 wants the second approach.
This also explains why the “Claude vs. ChatGPT” race the author mentions keeps flipping. Each new model changes the rules a little. People who keep using last year’s prompts get last year’s results, or slower versions of them.
Quick tips to keep in mind
- Don’t swap one crutch for another. Replacing “think hard” with ALL CAPS or “this is VERY important” misses the point.
- Describe the difficulty, not the effort. If a task really is tricky, say why, for example “these two contract clauses seem to conflict.” That tells Claude where to look.
- Watch the speed. Faster first responses are a good sign your cleanup worked.
- Revisit old workflows. Automations and API prompts built for older models probably carry the same baggage.
I love tips like this because they take thirty seconds to apply and you feel the difference right away. Deleting a few words is about the lowest-effort upgrade you can make.
This is just the first step in the author’s breakdown. The full post goes further into how to prompt Opus 5.5 and which mistakes to avoid, so check out the original LinkedIn post for the rest.