The single most repeated piece of prompt engineering advice from 2023 is now actively degrading your outputs.

"Think step by step." Wei et al.'s 2022 chain-of-thought paper showed it could boost GSM8K accuracy from 17.9% to 58.1% on PaLM 540B. Every tutorial, every LinkedIn carousel, every "10x your AI" thread parroted it. And for non-reasoning models, it genuinely worked — you were doing the model's thinking for it, scaffolding a process it couldn't manage alone.

That scaffolding is now getting in the way.

The Models Already Think

GPT-5, Claude Opus 4.6, Gemini 3.1 — every frontier model shipping today has internalized chain-of-thought. They reason before they respond. OpenAI's o-series models were the first to make this explicit: built-in multi-step reasoning that fires whether you ask for it or not. Claude followed with extended thinking, now evolved into adaptive thinking that calibrates depth to problem complexity automatically.

When you prepend "think step by step" to a query aimed at one of these models, you're not helping. You're prescribing a reasoning path that competes with the model's own, often superior, internal process. The Anthropic docs put it bluntly: "A prompt like 'think thoroughly' often produces better reasoning than a hand-written step-by-step plan. Claude's reasoning frequently exceeds what a human would prescribe."

PromptHub tested this directly and found that explicit chain-of-thought instructions on reasoning models "redirect the model's internal reasoning instead of guiding it." You're not adding a boost. You're adding interference.

Five Techniques That Flipped From Help to Harm

Chain-of-thought isn't the only casualty. Here's what else stopped working when models learned to reason:

Few-shot examples for reasoning tasks. They used to teach the model how to think. Now they constrain it. Few-shot still works for format alignment — showing the model what your output should look like — but feeding it worked examples of reasoning just narrows the solution space. The model has seen more math proofs, more code traces, more logical deductions than your three examples could ever capture.

Self-consistency prompting. The idea was to sample multiple reasoning paths and take the majority vote. But reasoning models are already dramatically more consistent by design. You're burning tokens on redundant generations that converge to the same answer.

Least-to-most and skeleton-of-thought. Both prescribe decomposition sequences. The model handles decomposition better when you let it choose its own structure. Telling Claude "first identify the subproblems, then solve each one" is like telling a surgeon "first pick up the scalpel, then make the incision." They know.

ALL CAPS and aggressive formatting. "YOU MUST", "CRITICAL:", "NEVER EVER" — these were hacks for models that needed strong signals to follow instructions reliably. Anthropic's migration guide now warns that Claude 4.5 and 4.6 will overtrigger on emphatic language. Where you once wrote CRITICAL: You MUST use this tool when..., you now just write Use this tool when.... The volume doesn't help. It distorts.

What Actually Works Now

The pattern is simple: define the destination, not the route.

Give the model a clear goal, tight constraints, and relevant context. Then get out of the way.

Bad (2023 playbook):
"Think step by step. First, identify the key variables.
Then, consider the relationships between them.
Next, formulate a hypothesis. Finally, test it
against the data and provide your conclusion."

Good (2026 reality):
"Determine whether this dataset shows evidence of
seasonal pricing effects. Use only the columns listed
below. State your confidence level and the two
strongest pieces of supporting evidence."

The first prompt micromanages the reasoning process. The second defines what success looks like and what materials to use. One competes with the model. The other collaborates with it.

The shift people are calling "context engineering" captures this well: the quality of what you feed the model matters more than how cleverly you phrase your request. The model is a reasoning engine now. Your job isn't to script its thinking — it's to load the right information into its context window. The metaphor that keeps circulating: the LLM is a CPU, the context window is RAM, and you are the operating system deciding what gets loaded.

This means the high-leverage work has moved upstream. Instead of tweaking prompt phrasing, you're curating which documents get included. Instead of adding more instructions, you're trimming context to the 150-300 word sweet spot where performance peaks (yes — performance degrades around 3,000 tokens of instructions, per recent testing). Instead of few-shot examples showing reasoning, you're providing reference data that grounds the model's own reasoning.

The Uncomfortable Part

Prompt engineering as most people learned it — the craft of phrasing — is shrinking in relevance. Not dead, but diminished. Role prompting still works. Clear constraints still matter. XML structure for complex inputs remains powerful. But the bag of tricks that defined the field in 2023 (chain-of-thought, few-shot reasoning, emphatic directives, multi-step scaffolding) has been absorbed by the models themselves.

The practitioners who adapt will focus on context architecture: what information does the model need, in what format, at what point in the conversation? The ones who keep tweaking "think step by step" variations will keep getting worse results and not understand why.

Anthropic's Claude documentation now includes a single line that should be taped above every prompt engineer's monitor:

"Prefer general instructions over prescriptive steps."

The models grew up. Our prompts need to catch up.