Drop "Think Step by Step": The Prompt Cut List, Backed by Evidence

What this is about

"You are a seasoned expert in …", "Think step by step", "IMPORTANT: ALWAYS follow this rule": phrases like these show up in German prompt guides, in company templates and in many AGENTS.md and CLAUDE.md files. Those are the text files you use to give coding agents standing instructions for a project. Back in 2024, PC-WELT's very first tip was to give the AI an expert role, because the word "expert" supposedly sends a "strong signal".

Within just a few weeks, Anthropic and OpenAI have both published guides that boil down to the same message: many of these habits do nothing for current models, or even slow them down. Our take: nobody had cleanly demonstrated a benefit for most of these tricks in the first place. What's new is that the vendors are now saying so themselves. Below, we put the vendor recommendations side by side with independent measurements, trick by trick.

What the vendors are now saying

Anthropic published a guide on November 12, 2026, on using Opus 5.5 as your daily driver for knowledge work. The Decoder and DiarioBitcoin covered it. The key points: delete "think hard" and "think carefully", because Opus 5.5 always thinks anyway. Start at the Medium effort level (the dial that sets how much compute the model spends before answering) and only move up to High if you need to. Ask for the finished document directly instead of an outline first. The developer docs, Prompting Claude Opus 5.5, go even further. If you want less thinking, lower the effort setting, because that reduces thinking "more reliably than prompt instructions". Anthropic also claims that Opus 5.5 on Medium matches or beats Opus 5 on High. Back in July, Anthropic wrote that it had cut more than 80 percent of Claude Code's system prompt "with no measurable loss" on its internal coding evals. The description of the to-do list tool alone shrank from about 9,100 characters to a few lines.

OpenAI makes a similar case for GPT-5.6 and GPT-6 Astra. The GPT-5.6 prompt guidance advises taking a working prompt, removing instructions one group at a time and re-testing after each step. Absolute words like ALWAYS, NEVER or "must" should be reserved for true invariants, such as safety rules. In the blog post "Rethinking skills and prompts for GPT-6 Astra", Eric Provencher writes about old, strict approval rules:

"Astra may take them too seriously and stop work at points where you'd actually be glad if it kept going."

Google breaks the pattern. For Gemini 4, the docs mostly tell you which parameters to remove: temperature, top_p and top_k are ignored. Instead of a thinking budget, there's now a thinking level (thinking_level). But the general prompting strategies page still says: "We recommend to always include few-shot examples in your prompts." Few-shot means giving the model a handful of worked examples. The page also recommends a role in the system prompt and considers "Think very hard before answering" helpful for difficult tasks. We found no dedicated Gemini 4 cut list.

The cut list at a glance

What independent studies measure

The most thorough series of measurements is the "Prompting Science Reports" from Wharton Generative AI Labs (Meincke, Mollick, Mollick, Shapiro et al.). They ask each question many times, because a single run tells you little with language models.

• Chain of thought (Report 2, June 2025) – the name for asking a model to think step by step. The test used 198 PhD-level GPQA Diamond questions, each run 25 times. With classic models, the trick helped on average: Gemini 2.0 Flash +13.5 percent, Claude 3.5 Sonnet +11.7 percent. In exchange, responses took 35 to 600 percent longer, and results became less consistent. With reasoning models, which think on their own before answering, almost nothing was left: o3-mini +2.9 percent, o4-mini +3.1 percent, Gemini 2.5 Flash −3.3 percent. The time penalty was 20 to 80 percent.
• Threats and tips (Report 3, August 2025). The researchers tested prompts like "I'll kick a puppy", a $1,000 tip or a trillion dollars. The trigger was Google co-founder Sergey Brin's claim that models perform better when threatened. Result: generally no significant effect. The one outlier, a gain of around ten percentage points for Gemini 2.0 Flash with a story about a mother with cancer, the authors chalk up to a quirk of that one model.
• Expert personas (Report 4, December 2025). Six models, 25 runs per question. No expert persona reliably improved GPQA accuracy. Roles like "toddler" or "layperson" often lowered it. When the expert role didn't match the subject, Gemini 2.5 Flash refused an average of 10.56 out of 25 answers, claiming it lacked the expertise.
• Politeness (Report 1, March 2025). Writing "Please" versus "I order" shifted individual questions by up to 60 percentage points. Across the whole dataset, though, it largely evened out.

The same pattern runs through all of them: on average, the trick does little, but on individual questions it can change a lot, in either direction. Report 3 found swings of +36 and −35 percentage points. So if you try a trick on three examples and see an improvement, you may just have measured noise.

On context files, two studies contradict each other. Researchers at ETH Zurich (arXiv 2602.11988, ICLR 2026) found that AGENTS.md files generally don't raise success rates but increase costs by more than 20 percent. Repository overviews provided no benefit, despite vendors recommending them. Lulla et al. (arXiv 2601.20404), on the other hand, measured a 28.6 percent shorter median run time and 16.6 percent fewer output tokens across 124 pull requests, with comparable success. The sample is small, though, and the authors only claim a correlation. Our take: it depends on what's in the file. A little, but unusual, information helps; repeating existing docs just costs you. That fits with the IFScale benchmark, where even the best models followed only 68 percent of instructions when given 500 at once.

Where the old tricks do work

The tricks aren't completely useless:

• Emphasis for format rules: A March 2026 study (arXiv 2603.23530) shows that models follow format instructions 2 to 21 percent less reliably when they're also busy with a task. For instructions about the end of the response, it's up to 50 percent. An explicit marker ("IMPORTANT") plus a reminder at the end restored 90 to 100 percent compliance in many settings. So IMPORTANT can help, if you use it once and deliberately.
• Roles for subjective tasks: In hate speech detection, persona prompts improved classification (arXiv 2601.20757). The model's explanations got worse, though. The Wharton authors also stress that they only measured accuracy, not tone or perspective.
• Examples for format: In Report 1, dropping format instructions lowered performance for both models tested.

The vendors aren't consistent either. Anthropic's own best-practices page still shows system prompts like "You are a helpful coding assistant specializing in Python". A design example there ends with "it is critical that you think outside the box!", even though the Opus 5.5 page explains that generic requests like "avoid a generic AI look" usually just swap one default style for another.

What works instead

Here's what all the guides agree on:

1. Define the goal and what done looks like. OpenAI recommends spelling out what "done" means, for example: implement, run, verify, fix errors. Requiring a check-in after the first draft, on the other hand, makes Astra stop too early.
2. Context, not incantations. Attach files and diagrams instead of retyping them, provide your house style and templates, and name unwanted patterns concretely.
3. Effort instead of "think carefully". According to Anthropic, the dial is the most important lever for quality, latency and cost. An independent test by David Krcek (rotecodefraktion.de) backs this up: Opus 5.5 on Medium solved all 30 runs, just like Opus 5 on High, but was 44 percent cheaper and about twice as fast at the median.
4. Grant permission explicitly. Instead of piling on prohibitions, approve safe workflows, such as "rerun affected tests without asking". Hard limits on risky actions still stay in place.

Our take

The guides retract less than it seems. "Think step by step" was a sensible trick for models that didn't think on their own. Now that thinking is built in and controlled by a dial, the line is dead weight. Expert personas, threats and tips, on the other hand, never had solid evidence behind them, only anecdotes.

Two caveats apply. The Wharton measurements come from 2024 and 2025 models and only measure multiple-choice accuracy. For Opus 5.5, GPT-6 Astra or Gemini 4, there are so far only vendor claims and a handful of individual tests. And the vendors have a vested interest: shorter prompts make their models look faster and cheaper.

Practical advice if you're switching models right now: don't carry over old prompts and skill files without checking them. Cut one group of rules at a time and re-measure on real tasks, with multiple runs.