Most people who use AI every day still open their prompts with four words that used to be magic: "think step by step." On the models you are using in August 2026, those four words are at best doing nothing, and at worst making your output measurably worse.
This is not a style preference. It is a consequence of an architecture change that shipped quietly across every major model family in the last twelve months, and almost nobody outside the API world noticed.
Here is what changed, which prompt habits to retire, and the exact template to use instead.
What is a reasoning model, and what changed in 2026?
A reasoning model is a language model that generates an internal chain of thought before it writes its visible answer. You do not see those tokens, but you pay for them and they drive the quality of the response. As of 2026 this behaviour is the default, not an option you switch on.
The clearest example is Claude Sonnet 5. According to Anthropic's adaptive thinking documentation, a request sent with no thinking field now runs with adaptive thinking enabled. On Claude Sonnet 4.6, that identical request ran with no thinking at all.
Anthropic's Sonnet 5 release notes go further: manual extended thinking now returns a 400 error, and setting temperature, top_p or top_k to non-default values also returns a 400 error. The dials you used to reach for have been removed.
OpenAI moved in the same direction. Reasoning depth on the GPT-5.6 family is set with a reasoning effort parameter, not with prose instructions inside your prompt.
Why does "think step by step" now hurt your output?
Telling a reasoning model to think step by step asks it to do something it has already done. The instruction lands in one of two bad places: the model narrates its reasoning into the visible answer, or it constrains the internal reasoning pass to follow your script instead of its own.
The first failure is the one you will actually notice. You ask for a 200-word product description, and you get 700 words of "First, let me consider the target audience. Second, let me think about tone." That is reasoning you already paid for, billed to you a second time as output.
The second failure is quieter and more expensive. Your enumerated steps are a ceiling. A model working through a pricing analysis may want to check five things; your prompt told it to check three, so it checks three.
There is a real-world version of this that practitioners hit constantly. You copy a prompt template written in 2023, paste it into a 2026 model, get a bloated response, and conclude the model got worse. The model did not get worse. Your prompt is fighting it.
Which five prompting habits should you retire in 2026?
Five techniques that genuinely improved output on 2023-era models now produce neutral or negative results on reasoning models. Each one was correct advice at the time. Each one is now overhead you are paying for in tokens, latency and drift.
1. "Think step by step" and its variants. This includes "reason through this carefully," "take your time," and "think harder." OpenAI's own guidance for current reasoning models is that you do not need to ask the model to think harder or to expose its chain of thought.
2. Long enumerated do's and don'ts. A 40-line list of rules was how you got consistency out of a non-reasoning model. On a reasoning model it competes with the task itself for attention, and every rule you add dilutes the ones that matter.
3. Cranking temperature for creativity. On Claude Sonnet 5 this is not a bad idea, it is an error. Non-default sampling parameters return a 400. Creative variation now comes from what you ask for, not from a slider.
4. "Be concise" as a standalone instruction. It is vague and it is the wrong lever. "Be concise" tells the model to compress everything equally. "Maximum 150 words, no preamble, no summary paragraph" tells it exactly where to cut.
5. Role prompting as the whole strategy. "You are a world-class marketing strategist" was load-bearing in 2023. In 2026 it is a mild framing nudge. The work has moved into context and constraints: what the model can see, and what counts as done.
What should you write instead? The outcome-first prompt template
Replace process instructions with outcome specification. A reasoning model does not need to be told how to think; it needs to be told what it is looking at, what it is optimising for, and what a finished answer looks like. Six blocks cover it: goal, context, constraints, evidence, success criteria, output format.
Here is a complete, copy-paste-ready prompt. It contains no reasoning instructions at all.
Try this prompt:
GOAL
Write the launch announcement email for our new appointment-booking feature.
CONTEXT
Audience: 2,400 existing customers, mostly clinic and salon owners in Hong Kong, average 4 staff.
They already use our scheduling product daily. They have asked for online booking for two years.
Our brand voice is direct and practical. We do not use exclamation marks.
Product facts: customers share one booking link, clients book without creating an account, bookings appear in the existing calendar within 5 seconds, included in all paid plans at no extra cost.
CONSTRAINTS
Maximum 180 words in the body. One subject line, maximum 45 characters.
No preamble before the email. No explanation after it.
Do not claim time savings we have not measured.
EVIDENCE
Every product claim in the email must map to a fact listed under CONTEXT. If you need a fact I have not given you, write [NEED: what you need] inline instead of inventing it.
SUCCESS CRITERIA
A clinic owner who skims only the subject line and first sentence understands that online booking now exists and costs nothing extra.
OUTPUT FORMAT
Subject line on the first line. Blank line. Email body. Nothing else.
Notice what is absent. No persona. No "think carefully." No numbered method. The model decides how to get there; you decide what "there" means.
The EVIDENCE block is the single highest-leverage line in that template. Telling a model to write [NEED: ...] rather than fill a gap converts a hallucination into a visible to-do item.
How do you apply this to a recurring work task?
The template pays off most on tasks you repeat weekly, because a stable prompt makes output variation diagnosable. When results wobble, the cause is now a change in your context block, not a mystery.
Take a weekly performance summary. The 2023 approach was a long instruction: act as a data analyst, review the figures step by step, first identify trends, then explain them, then recommend actions, be concise but thorough.
The 2026 approach fixes GOAL, CONSTRAINTS, SUCCESS CRITERIA and OUTPUT FORMAT permanently in a saved prompt, and changes only CONTEXT each week when you paste in the new numbers.
Success criteria are where most people underinvest. "Write a summary of this week's numbers" has no finish line. "A manager who reads only the first three lines can name the one metric that moved and the one decision required this week" has a finish line the model can aim at, and that you can grade against.
This is also why a saved prompt beats a clever prompt. A prompt you reuse 40 times and refine 6 times will outperform a brilliant one-off, because you can see which edit changed which behaviour.
Where does this advice break down?
Dropping reasoning instructions is correct for current frontier reasoning models. It is not universally correct. Three situations still justify the old habits, and pretending otherwise would leave you worse off.
Small and older models. If you are running a lightweight local model or an older non-reasoning model, chain-of-thought prompting still works, and it works well. The advice in this article is scoped to models that reason by default.
When you need to audit the logic. Sometimes you genuinely want the reasoning in the visible output because you are checking it, for example a compliance calculation or a pricing decision that someone will question. In that case ask for it as a deliverable: "After the answer, list the assumptions you used." That is an output requirement, not a reasoning instruction, and the difference matters.
Structured extraction and formatting jobs. Pulling fields out of 300 invoices does not benefit from reasoning at all. On Claude Sonnet 5 you can pass a disabled thinking setting to switch it off, which cuts both cost and latency. Choosing the cheapest adequate mode is the same discipline as choosing the cheapest adequate model tier, which we covered in our guide to choosing between GPT-5.6 Luna, Terra and Sol.
One more caution. Model behaviour is version-specific. The 400-error behaviour described here is documented for Claude Sonnet 5 as of August 2026. Before you rebuild a production workflow on any of this, check the current release notes for the exact model string you are calling.
Try it now: the 20-minute prompt audit
You can test every claim in this article in about twenty minutes using prompts you already have. The point is not to trust the argument; it is to see the difference in your own outputs, on your own tasks.
Open the three prompts you reuse most. For each one, do this:
--- Delete every instruction that tells the model how to think. That includes "step by step," "carefully," "as an expert," and any numbered method.
--- Add a CONSTRAINTS block with a hard word limit and an explicit "no preamble" rule.
--- Add a SUCCESS CRITERIA line describing what a reader should be able to do after reading the output.
--- Run the original and the rewritten version on the same input, in two separate chats, and compare length, accuracy and how much editing each needs.
In our own testing the rewritten prompts produce shorter output that needs fewer edits, and the gap is widest on tasks with a clear deliverable such as emails, briefs and summaries. Open-ended brainstorming shows the smallest difference.
Keep whichever version wins. That is the whole method, and it is worth repeating every time a model you depend on ships a major version.
The takeaway: stop managing the model, start specifying the outcome
The skill that mattered in 2023 was knowing how to make a model think. The skill that matters in 2026 is knowing how to describe a finished piece of work precisely enough that a model which already thinks can hit it.
That is a genuinely different skill, and it is closer to good briefing than to programming. If you have ever written a clear brief for a freelancer, you already have most of it.
It also explains a frustration a lot of capable people feel right now. You are not behind because you lack a secret technique. You are behind because the techniques you learned were correct for a generation of models that has quietly been replaced.
Keeping up with that kind of change is exactly where a partner helps. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Put Your AI Skills to the Test
Rewriting your prompts is step one. Turning them into a workflow that runs reliably every week is where the real gain sits. We'll walk you through every step, from prompt design and model selection to deployment inside your day-to-day tools.
Reviewed by the UD AI team. Model behaviour cited in this article reflects vendor documentation published as of 6 August 2026 and is version-specific.
