Felix Laurent
Learning how to prompt Claude in 2026 means unlearning a few old tricks. The newest models, Claude Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.2, follow instructions more literally than their predecessors. Prompts full of capital letters and "CRITICAL" warnings now push them too far. Two popular tools are also gone: prefilled responses stopped working with the Claude 4.6 models, and the Console prompt generator was retired in August 2026.
This guide covers the techniques Anthropic recommends for its current models and what changed with Opus 5.5, Sonnet 5.5 and Haiku 5.5. It also shows easier ways to get a good prompt without writing every line yourself. Everything was checked against Anthropic's documentation and help center on November 12, 2026.
What makes prompting Claude different in 2026
Anthropic's own description of the reader is the best place to start. Its prompting docs ask you to think of Claude as "a brilliant but new employee who lacks context on your norms and workflows". The test that follows is simple. If a colleague with little context would be confused by your prompt, Claude will be too.
Three things changed for the current models. They take each word at face value, so emphasis that once kept older models on track now makes them overreact. They decide on their own how long to think, and a separate effort setting controls how much. And the "add more detail" instinct backfires: Anthropic says prompts and skills written for earlier models "are often too prescriptive" and "can degrade output quality".
So good Claude prompts in 2026 are shorter on rules and longer on context. The core techniques below show how to strike that balance.
The core techniques for writing Claude prompts
Anthropic's prompting best practices page is the living reference for every current model. It lists the techniques in a set order, from the most basic to the most specialized, and the sections below follow it.
Be clear and direct
Say exactly what you want, including what a good answer looks like. Claude does not guess at ambition you did not state. If you want more than the minimum, Anthropic advises you to "explicitly request it rather than relying on the model to infer". Lead with the task, name the audience, and set the length, tone and format you need.
Add context and the reason behind each rule
A rule with a reason works better than a rule alone. Anthropic's example is a prompt that says "NEVER use ellipses". It works better when the prompt explains that a text-to-speech engine will read the answer aloud. The docs note that "Claude can generalize from the explanation", so it also avoids other things a speech engine would stumble over.
Use three to five examples
Anthropic calls examples "one of the most reliable ways to steer Claude's output format, tone, and structure". Its docs recommend three to five examples for the best results. Make them relevant to the real task, vary them so Claude does not copy one pattern, and wrap each in <example> tags. You can also ask Claude to judge your examples or generate more of them.
Structure complex prompts with XML tags
When a prompt mixes instructions, background and input data, XML tags keep the parts apart. Use tags such as <instructions>, <context> and <input>, keep the names consistent, and nest them where the content has a natural hierarchy. Claude has no special list of tag names, so pick names that describe what is inside.
Give Claude a role
A role sets the expertise and tone for a whole conversation. Put it in the system prompt, or at the top of a chat. Anthropic says "even a single sentence makes a difference", for example "You are a senior tax adviser reviewing a small business's records".
Put long documents first
For prompts over about 20,000 tokens, place the documents at the top and the question at the end. Anthropic reports that "queries at the end can improve response quality by up to 30 percent". Wrap each document in <document> tags with its source. For long reports, ask Claude to quote the relevant passages first and then answer from those quotes.
Control the output format
Describe the format you want instead of listing what to avoid. "Write in flowing prose paragraphs" works better than "don't use markdown". The style of your prompt also leaks into the answer, so a prompt without markdown tends to get a reply without it. Anthropic notes the newest models are less verbose and may skip detailed summaries unless you ask for one.
Ask Claude to check its work
Accuracy improves when the prompt tells Claude how to handle gaps in the information you provide. Give it permission to say it does not know, and ask it to flag any claim your documents do not support. For key figures, ask Claude to quote the line each number comes from.
A closing line such as "Before you finish, check your answer against the criteria above" adds a cheap self-review. It gives Claude a chance to catch its own mistakes before you read the answer, and it saves a round of corrections.
Say whether Claude should act or advise
With tools, the wording decides what happens. "Can you suggest some changes" gets suggestions. "Change this function" gets the edit. If you only want ideas, say so and tell Claude to stop there.
These techniques apply to every current model. Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.2 add a few rules of their own.
What changed with Opus 5.5 and Fable 5.2
Most changes come from one trait: the new models follow instructions more precisely. Prompts written to push older, less attentive models now overshoot. Anthropic's model-specific prompting pages list the adjustments.
- Soften aggressive language. Anthropic says capitalized warnings "might now overtrigger". Write "Use this tool when…" instead of "CRITICAL: You MUST use this tool when…".
- Treat "never" with care. Fable 5.2 reads any rule with "never" as absolute. "Ask for approval before deleting files" leaves room that "never delete files" does not.
- Drop "think carefully" lines in chat. On Opus 5.5, removing such a line made replies start sooner with no clear loss in quality.
- Limit scope on Sonnet 5.5. It tends to add tests, docs and small files nobody asked for. Anthropic suggests: "Don't add features, tests, files, docs, or refactors that weren't asked for."
- Name the patterns to avoid. For design work on Opus 5.5, "avoid a generic AI look" just swaps one default style for another. List the specific patterns you don't want instead.
- Give Haiku 5.5 today's date when it has a search tool, and avoid blanket rules such as "search for any present-day question".
- Revisit old style rules. Fable 5.2 obeys them closely, so a hard limit such as 20 words per sentence can make its prose choppy.
Approvals in long tasks need the same precision. Anthropic advises saying exactly what you want next instead of sending a bare "continue". How much the model thinks before it answers is now a setting of its own.
Thinking, effort and prompt chaining
Classic chain-of-thought prompting asked the model to "think step by step" in its answer. Current Claude models do that internally with adaptive thinking, deciding when and how much to think. On Opus 5.5 and the Fable models, thinking is always on.
The effort parameter replaced most "think more" or "think less" instructions. Opus 5.5 and Haiku 5.5 default to medium, while Sonnet 5.5 and Fable 5.2 default to high. Anthropic's advice is direct: "To get less thinking, lower the effort level". Effort barely changes answer length, so set the length in the prompt.
Prompt chaining, splitting a task into several prompts, is now mostly optional. Anthropic says "Claude handles most multistep reasoning internally". A chain still helps for self-correction: one prompt drafts, a second reviews the draft against your criteria, and a third refines it. All of this assumes you write the prompt yourself, and often you don't have to.
Easier ways to get a good Claude prompt
Several features save you from writing a perfect prompt by hand. One popular tool from 2024 and 2025 is gone, so it is worth knowing what still works.
Ask Claude to write or improve the prompt
The simplest prompt generator is Claude itself. Describe the task, the audience and a sample of the input, then ask Claude to write a prompt for it. Paste an existing prompt and ask what is unclear, missing or contradictory. Anthropic's prompting overview links a "prompt generator notebook", a metaprompt recipe in the Claude Cookbook that does this for developers.
You can also ask Claude to interview you first. A request such as "Ask me five questions about this task before you write anything" surfaces the context you forgot to give. It is a quick way to turn a vague idea into a complete brief.
Note: the Console prompt generator is gone
Many guides still point to the "Generate a prompt" and prompt improver buttons in the Claude Console. Anthropic retired the legacy Workbench on August 17, 2026, along with saved prompts, variables and the evaluation tool. Its replacement, the playground, supports every Messages API parameter but does not generate, save or evaluate prompts. The old prompt library page now redirects to the best practices docs.
Save instructions once with Projects, Instructions and memory
Repeating the same context in every chat is the most common waste of effort. In claude.ai, Projects are workspaces with their own chat history, knowledge files and instructions, and free accounts can create up to five. Instructions for Claude, found under Settings, apply to every chat on the account.
Memory works in the background. Claude saves topics as you chat, and you can view, edit or delete them under Settings. Memory is on by default on Free, Pro, Max and Ultra plans, and each project keeps its own.
Turn repeated tasks into Skills
A skill is a folder of instructions, scripts and resources that Claude loads only when a task needs it. Anthropic ships skills for Excel, Word, PowerPoint and PDF files, and you can write your own. They are available on every plan, including Free, with code execution turned on. A prompt you paste every week is a good candidate for a skill.
Learn from Anthropic's free tutorial
Anthropic keeps an interactive prompt engineering tutorial on GitHub, with nine chapters and an appendix of exercises. It was written for Claude 3 models, and Anthropic's overview says to follow the current best practices where the two differ. Anthropic Academy links the same tutorial alongside courses on building with the Claude API and on evals. For a beginner, it is the most effective way to practice each technique on real tasks.
Use CLAUDE.md and plan mode in Claude Code
In Claude Code, a CLAUDE.md file holds the instructions for a project. Claude reads it at the start of every session, and Anthropic suggests keeping it under about 200 lines. Run /init to generate a first version from the codebase. Plan mode, started with Shift+Tab, makes Claude propose a plan before it edits anything. Anthropic's tip for coding prompts is to "state the outcome, not the steps".
These tools cut down on prompt writing, but a good prompt still has the same shape. An example shows it best.
A Claude prompt example: before and after
Here is a weak prompt a person might send to Claude:
Summarize this report and make it good.Claude has to guess the reader, the length, the format and what "good" means. A full prompt answers those questions:
You are a financial analyst writing for our board of directors.
<context>
The board meets on Friday and has 10 minutes for this item.
They care about cash runway and hiring, not product details.
</context>
<instructions>
Summarize the report below in five bullet points, under 120 words in total.
Lead with the runway figure. Flag any number the report does not support.
</instructions>
<document source="Q3 operations report">
{{REPORT}}
</document>The full version sets a role, explains why brevity matters, separates the parts with XML tags and names the format. It uses no capital letters, no warnings and no rules without reasons. Most weak prompts fail in a handful of predictable ways.
Common Claude prompting mistakes
The same mistakes show up in most prompts that disappoint. Each has a simple fix.
- A vague request. "Make it better" gives Claude nothing to aim at. State the goal and what a good answer looks like.
- Rules without reasons. Explain why each rule exists so Claude can apply it to cases you did not list.
- Only negative instructions. Say what to do. A list of things to avoid leaves the actual target open.
- Shouting. Capital letters and "MUST" make current models overreact. Plain sentences work better.
- Prefilling the answer. Anthropic stopped supporting prefilled responses with the Claude 4.6 models, and such API requests return an error. Ask for the format in the prompt, or use structured outputs.
- Old prompts carried over. Instructions written for Claude 3 or 4 are often too heavy for Opus 5.5 and the other current models. Cut them back and test again.
- Starting over instead of iterating. Tell Claude what was wrong with the last answer. Each correction adds context the first prompt lacked.
Avoiding these mistakes gets most people further than any template. That leaves the question this guide set out to answer.
So how should you prompt Claude?
Prompt Claude the way you would brief a sharp new colleague. Give the goal, the audience and the reason behind each rule. Add three to five examples when the format matters, and use XML tags for prompts with several parts. Leave out the capital letters and the "think step by step" lines, and use the effort setting to change how hard the model thinks.
For anything you do more than once, stop rewriting the prompt. Put standing instructions in a Project, in Instructions for Claude, in a Skill or in CLAUDE.md. When you are unsure what to write, describe the task and ask Claude to write the prompt or interview you first. For general prompting advice that applies to every AI model, see our guide on how to write AI prompts.
FAQ
The questions below come up most often about prompting Claude.
Is prompting Claude different from prompting ChatGPT?
The basics are the same: be clear, give context and show examples. Anthropic puts more weight on XML tags for structure, and current Claude models take instructions literally. Emphasis and "never" rules that other models need can make Claude overcautious.
How long should a Claude prompt be?
As long as the context requires and no longer. A short task needs a few sentences. A complex one can run to pages, provided the documents come first and the question comes last.
Should I use a system prompt or a user prompt?
Use the system prompt for the role and the rules that apply to the whole conversation. Put the task and its data in the user message. In claude.ai, Project instructions and Instructions for Claude work like a system prompt.
Is there an official Claude prompt generator?
Not in the Console anymore. Anthropic retired the prompt generator and prompt improver with the Workbench on August 17, 2026. Developers can still use the metaprompt notebook in the Claude Cookbook, and anyone can ask Claude to write or improve a prompt.
Can I save and reuse Claude prompts?
Yes. In claude.ai, use Projects, Instructions for Claude or a custom Skill. In Claude Code, put the instructions in CLAUDE.md or turn them into a skill you call with a slash command.
