Prompt Engineering 2026 Prompting Manual: Architecting High-Performance AI Outputs

Start hereIn 2022, adding “Let's think step by step” lifted one model's arithmetic score from 17.7% to 78.7%. Today OpenAI says the same instruction is unnecessary for its reasoning models. So which prompting habits still help?

01A template to adapt

Copy the template below, change the parts in brackets, and test it on the model you use. It separates five parts of a prompt: who the model should be, the source material, the tasks, the limits and the output format. The five block names are this page’s own labels, not a standard. Any clear, consistent labels do the same job.

### SYSTEM PERSONA
You are a senior systems architect. Write directly and clearly, without introductions or pleasantries.

### INPUT DATA
<technical_specification>
[Paste your code, notes or documentation here]
</technical_specification>

### TASKS
1. Read the <technical_specification> block and identify the three most serious bottlenecks.
2. For each bottleneck, outline one concrete fix.
3. Use web search to check whether any recent security guidance affects these fixes, and name the source for each point.

### GUARDRAILS
- No introductory phrases, filler or closing summaries.
- Base every statement on the <technical_specification> block or on sources you found with web search, and name them.
- If the material is not enough to complete a task, write "Data Insufficient" instead of guessing.

### OUTPUT FORMAT
Markdown, with ### headings for each part. Put the findings in a three-column table:

| Bottleneck | Fix | Relevant guidance and source |
| :--- | :--- | :--- |

The steps below explain each part, and where the advice comes from.

PersonaExpertise, tone and limits

Input dataThe source material, inside tags

TasksPlain, numbered instructions

GuardrailsWhat the answer may not do, and when to use tools

Output formatThe exact shape of the answer

The block names are this page's own labels, not a standard.

02Step by step

1. Mark out the parts of the prompt

The three labs’ guides agree on this more than on anything else. Separate instructions from material, so the model can tell them apart.

• Anthropic: XML tags help Claude read complex prompts without ambiguity, especially when instructions, context, examples and variable inputs are mixed.
• OpenAI: use delimiters such as Markdown, XML tags and section titles to mark the distinct parts of the input.
• Google: XML-style tags such as <context> and <task> work, and so do Markdown headings. Pick one format and use it consistently within a single prompt.

2. Put long documents above the question

For long inputs, around 20,000 tokens or more, Anthropic suggests placing the documents near the top of the prompt, above the question and instructions. It reports that questions at the end can improve response quality by up to 30 percent in tests, especially with complex inputs made of several documents.

3. Write plain, direct instructions

The labs’ guides ask for clear, direct instructions rather than any particular tone. Emphasis in capitals can now backfire. Anthropic’s guide says Claude Opus 4.5 and 4.6 respond more closely to the system prompt than earlier models. So wording such as CRITICAL: you MUST use this tool when… can make them use a tool too often. It suggests plain wording such as Use this tool when… instead. Google’s guide says to avoid unnecessary or overly persuasive language.

Reasoning models no longer need step-by-step phrasing. OpenAI’s guide says prompting its reasoning models to think step by step or explain their reasoning is unnecessary. It adds that such instructions may not help, and can sometimes hinder. Anthropic’s guide says a general instruction such as think thoroughly often produces better reasoning than a hand-written step-by-step plan.

4. Let the model say it does not know

The Data Insufficient line in the template follows advice in Anthropic’s guide on reducing hallucinations: give the model explicit permission to say it does not know. The same guide says its techniques significantly reduce made-up answers but do not eliminate them, and advises checking critical information. For more on why models invent answers, see Why AI makes things up.

5. Make the model use a tool for calculations

Anthropic, OpenAI and Google each offer a code-execution tool. It lets the model write and run code in a separate environment; Anthropic and OpenAI describe theirs as a sandbox. Web search is offered as a tool too. OpenAI describes its code interpreter as a way to solve problems in data analysis, coding and maths by writing and running Python.

• Vague: Calculate the compound growth rate in this spreadsheet.
• Specific: Write and run a Python script to compute the compound growth rate from the attached data. Do not estimate the figure. Show the script and its output alongside the final number.

6. Re-ground a long conversation

In a long chat, earlier material can drift out of the model’s effective attention (see Why it works, below). One practical response is to ask the model to restate where things stand before a new phase of work:

Before we continue, review our conversation so far. Summarise the current status, the decisions we have made and the open questions in three bullet points. Write nothing else.

This is a working habit rather than a tested technique; we found no study of this exact prompt. The same idea is built into some products: Anthropic’s compaction feature automatically summarises earlier parts of the conversation so a long session can continue. Another option is to start a fresh conversation and paste the summary in.

7. Check the draft against something outside the model

Anthropic’s guide suggests adding a line such as Before you finish, verify your answer against [test criteria]. It notes that one of its models, Claude Opus 5, checks its own work well without being asked. On Opus 5, such instructions can cause over-verification and should be removed. On other models, a review prompt that gives the model something outside itself to check against looks like this:

Before giving your final answer, check your draft against [the test results / the source documents / the output of the code you ran]. List anything in the draft they do not support, revise the draft, then show only the revised version.

03Why it works

Three changes explain why older prompting habits matter less than they did, and one line of research explains step 7.

Models reason before they answer

The phrase Let’s think step by step comes from a 2022 paper by Takeshi Kojima and colleagues. Adding it before each answer raised one large model’s accuracy on the MultiArith arithmetic test from 17.7% to 78.7%. Reasoning models now do that work internally, which is why the labs’ advice in step 3 has changed.

Models can hold very long inputs

Anthropic lists a 1M-token context window for Claude Opus 5.5, the model it suggests starting with, and for each of its other current models. OpenAI lists 1.05M tokens for GPT-6 Astra, its flagship model, and Google lists an input limit of 1,048,576 tokens for Gemini 3.8 Flash. Google’s prompting guide puts 100 tokens at roughly 60 to 80 words, so a million tokens is somewhere around 600,000 to 800,000 words.

Longer is not automatically better. Anthropic’s documentation says that as the token count grows, accuracy and recall degrade, which it calls context rot. A 2023 study by Nelson Liu and colleagues, Lost in the Middle, found that models did best with information at the start or end of a long input. They did much worse with information in the middle.

Self-review works best with something to check against

Asking a model to critique and revise its own draft has mixed evidence. A 2023 paper, Self-Refine (Aman Madaan and colleagues), reported that outputs improved by about 20% in absolute terms on average across seven tasks. That was when a model gave itself feedback and revised. A later paper by Jie Huang and colleagues, presented at ICLR 2024, found that models struggle to correct their own reasoning without outside feedback. Their performance sometimes got worse after self-correction.

The two results fit together better when the review has something outside the model to check against. A 2024 survey of self-correction research by Ryo Kamoi and colleagues, in Transactions of the Association for Computational Linguistics, found that earlier studies often used unfair evaluations that overstated self-correction. It concluded that self-correction works well in tasks that can use reliable external feedback.

04Older habits and what the guides say now

Older habit	What current guides and research say
Adding Let’s think step by step to every question	It helped older models a great deal (Kojima and colleagues, 2022). OpenAI says it is unnecessary for its reasoning models and can sometimes hinder them.
Being very polite, or very rude, to get a better answer	The research is mixed. A 2024 study in English, Chinese and Japanese found that impolite prompts often did worse. A small 2025 study (50 questions, one model, ChatGPT-4o) found very rude prompts scored slightly higher: 84.8%, against 80.8% for very polite ones. The labs’ guides ask for clear, direct instructions rather than any particular tone.
Emphasis in capitals: CRITICAL: you MUST…	Anthropic suggests plain wording instead, and Google advises against overly persuasive language (step 3).
Cutting source material down to fit a small limit	The models named above accept around a million tokens, so the full text often fits. Choosing what goes in still matters, because accuracy and recall degrade as the context grows.
Taking the first answer as finished	A review step can help, especially when it checks the draft against something outside the model, such as test results, a tool’s output or sources.

Model behaviour changes within months. This page was written in June 2026 and last brought up to date in October 2026, so test anything here against the model you use.

Old	“Let's think step by step” on every question
Now	OpenAI: unnecessary for its reasoning models, can sometimes hinder
Old	“CRITICAL: you MUST use this tool when…”
Now	Anthropic: plain wording such as “Use this tool when…”
Old	Cut the source down to fit a small limit
Now	The full text often fits; choose what goes in
Old	Take the first answer as finished
Now	Check the draft against tests, tools or sources

An October 2026 snapshot of the labs' own guides.

05A simple test routine

Anthropic’s guide makes the same point about its own advice: treat a tip as measured on one model, and re-check it against your own tests before relying on it with another. A routine for doing that with your own prompts:

Step	What to do
1	Write a short list of real questions from your own work, with answers you already know.
2	Run them through your current prompt and note which answers are right.
3	Change one thing in the prompt, run the list again and compare.
4	Run the list again whenever the prompt or the model changes.

For the terms used on this page, see the AI glossary. For how these models work underneath, see How large language models work.

06Quick answers

07Use this

08Quick quiz

How we checked this · Last checked October 2026

• Prompting advice on this page comes only from the three labs' own documentation: Anthropic's prompting, context-window, model and hallucination pages, OpenAI's reasoning guide, models page and code interpreter page, and Google's prompt design, model and code execution pages.
• Model names, recommended starting models and context windows were reread on the labs' own pages on 13 October 2026. They change within months; the next recheck is due in November 2026.
• Research claims come from the papers themselves (arXiv abstract pages and their journal or conference records). The 2025 politeness study is a small preprint: 50 questions, one model.
• We found no study of the exact re-grounding prompt in step 6, and the page labels it a working habit. The five block names in the template are our own labels.

How we check every article →

09Sources