How to Prompt Claude Fable 5 Like an Anthropic Engineer: 6 Rules That Actually Work
Anthropic's own best practices for Claude Fable 5 include giving context, negative prompting, effort levels, and avoiding reasoning requests that trigger Opus.
Six Prompting Rules That Make Claude Actually Useful
Most people who complain that Claude gives vague, over-cautious, or verbose responses are prompting it the same way they’d type a Google search. That works fine for a search engine. It doesn’t work for a model like Claude Fable 5, which responds very differently depending on how much context, structure, and direction you give it.
Anthropic has published a substantial amount of guidance on how their engineers actually prompt Claude. These aren’t abstract principles — they’re specific behaviors built into how Claude interprets instructions. Understanding them means the difference between a model that hedges everything and one that produces sharp, usable output on the first try.
Here are six rules drawn directly from Anthropic’s documented best practices, with practical examples of each.
Rule 1: Give Claude a Role Before You Give It a Task
Claude responds to context about who it is before responding to what you want it to do. This is called role prompting, and it’s one of the most reliable ways to improve output quality consistently.
The mechanism is straightforward: Claude has internalized a huge amount of knowledge about how different professionals think, write, and reason. Activating that framing early shifts its entire approach to the task.
What this looks like in practice
A weak prompt: “Summarize this contract.”
A stronger prompt: “You are a contracts attorney reviewing an agreement for a startup founder. Summarize the key risks, obligations, and any unusual clauses in this contract.”
The second version isn’t just more detailed — it tells Claude which lens to use. The “contracts attorney” framing triggers a specific mode of reading that prioritizes risk identification over general comprehension.
How specific should the role be?
More specific is almost always better, but the role needs to be coherent. “You are a senior copywriter with 15 years of B2B SaaS experience” will outperform “you are a writer” for most content tasks. But “you are a futuristic AI philosopher” may produce results that sound interesting but aren’t practically useful.
Use roles that map to real expertise you’d actually hire for the task.
Rule 2: Use Negative Prompting to Remove Default Behaviors
Claude has default behaviors built in — tendencies toward certain formats, lengths, caveats, and tones. Some of these are useful. Many are not.
Negative prompting means explicitly telling Claude what not to do. This is one of the most underused prompting techniques, even though it directly addresses the most common frustrations people have with Claude’s output.
The defaults that trip people up most
- Hedging and caveats — Claude often adds “it’s worth noting that…” or “keep in mind that…” even when the user already knows the context.
- Preamble — Claude tends to restate the task before answering it, which wastes tokens and clutters output.
- Over-qualification — Phrases like “this may vary,” “depending on your situation,” or “I’d recommend consulting a professional” appear frequently, even when they add nothing.
- Padding — Claude sometimes adds a closing paragraph summarizing what it just said.
How to write effective negative prompts
Be direct. Say exactly what you want removed.
“Do not add caveats, disclaimers, or suggestions to consult a professional. Do not restate the task before answering. Do not include a summary paragraph at the end.”
You can include this as a standing instruction in your system prompt if you’re building an AI application. This way it applies to every response without needing to be repeated.
Negative prompts compound well with role prompts. Combining both gives you a model that knows who it is and what to leave out — which is usually enough to produce clean, direct output.
Rule 3: Specify the Effort Level You Actually Need
Claude doesn’t know whether you want a two-sentence answer or a 2,000-word analysis unless you tell it. Left to its own judgment, it will often produce something in the middle — too long for quick tasks, too shallow for deep ones.
Anthropic refers to this dimension as “effort level,” and it’s one of the most practical controls you have over Claude’s output.
Four effort levels worth knowing
Minimal — A sentence or two. Good for quick lookups, yes/no decisions, or simple definitions.
- “In one sentence, what is the difference between a revocable and irrevocable trust?”
Concise — A short paragraph. Good for explanations that need context but not elaboration.
- “In 2–3 sentences, explain why interest rates affect bond prices inversely.”
Standard — A full response with structure. Good for most content and analysis tasks.
- “Write a structured explanation of [X], covering the main concept, a practical example, and one common misconception.”
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
Comprehensive — Thorough, detailed, multi-part output. Good for research synthesis, long-form writing, or complex analysis.
- “Provide a comprehensive analysis of [X]. Include historical context, current best practices, key debates in the field, and practical recommendations.”
Why this matters for speed and cost
Claude Fable 5 calibrates its compute based on the apparent complexity of the task. A vague prompt that could require either a one-line answer or an essay creates ambiguity, and Claude tends to resolve that ambiguity by going longer. Being explicit about effort level saves time, reduces unnecessary verbosity, and keeps costs predictable when you’re running Claude at scale.
Rule 4: Avoid Triggering Extended Thinking When You Don’t Need It
This is the rule most people don’t know about — and it directly affects which version of Claude’s reasoning is applied to your request.
Claude Fable 5 has a built-in mechanism for detecting when a request requires complex, step-by-step reasoning. When it detects this, it can switch into an extended thinking mode — a deeper, slower, more compute-intensive process. In Anthropic’s model lineup, this is associated with Opus-tier processing.
Extended thinking is valuable for genuinely hard problems: multi-step math, complex logical deductions, intricate legal or medical reasoning. But it’s overkill for most tasks, and triggering it unnecessarily makes responses slower and more expensive.
What accidentally triggers extended thinking
- Asking Claude to “reason through” or “think step by step” about something simple
- Phrasing requests as puzzles or problems when they’re actually just information retrieval
- Adding “carefully consider all angles” or similar framing to routine tasks
- Asking for analysis of highly ambiguous or contradictory information
How to avoid it when you don’t need it
Keep your prompt direct and action-oriented. If you want a list, ask for a list. If you want a summary, ask for a summary. Don’t frame straightforward tasks as reasoning problems.
Instead of: “Think carefully about the pros and cons of remote work and reason through which factors matter most.”
Use: “List the five most significant advantages and disadvantages of remote work for knowledge workers.”
The second version is concrete and answerable without deep deliberation. It gets you a better-structured response faster.
When you do need extended thinking — for genuinely complex analysis — lean into it intentionally. Ask Claude to “reason step by step” or “work through this systematically.” That’s when the extra compute pays off.
Rule 5: Use XML Tags to Structure Inputs
Claude is trained to recognize and use XML tags as structural markers. This is one of Anthropic’s own internal practices, and it meaningfully improves output quality for any prompt that contains multiple distinct components.
Most prompts have at least two parts: instructions and content. If those two things are mixed together in plain prose, Claude has to infer where one ends and the other begins. Tags remove that ambiguity entirely.
A basic example
Without tags: “Here is the article I wrote about market trends. Please edit it for clarity and fix any grammatical errors. The article is about semiconductor supply chains and was written for a non-technical audience…”
With tags:
<task>Edit the following article for clarity and correct any grammatical errors. The audience is non-technical.</task>
<article>
[Your article text here]
</article>Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
The tag version is unambiguous. Claude knows exactly what the instructions are and what the content is. It doesn’t have to parse prose to separate them.
When tags matter most
Tags make the biggest difference in prompts that include:
- Long documents being passed as context
- Multiple distinct inputs (e.g., an email thread plus a brief)
- Examples you want Claude to learn from (put them in
<example>tags) - Sections you want Claude to treat differently
For short, simple prompts, tags are optional. For anything complex, they’re worth adding — they take ten seconds and consistently improve output quality.
Rule 6: Give Claude an Example, Not Just a Description
The most reliable way to get Claude to match a specific style, format, or tone is to show it what you want rather than describe it. This is called few-shot prompting, and it’s one of the most well-documented techniques in the prompt engineering literature.
Descriptions are inherently ambiguous. “Write in a conversational but professional tone” means something different to Claude than it does to you — and Claude’s interpretation may not match yours. An example removes that ambiguity.
How few-shot prompting works
Provide one or two examples of the output you want, clearly labeled, before asking for the actual output.
<instruction>Write a product update announcement for our newsletter. Match the tone and format of the example below.</instruction>
<example>
We just shipped keyboard shortcuts for the dashboard — something a lot of you have asked for. Press Cmd+K to open the search panel, and use arrow keys to navigate. More details in the docs. Thanks for the feedback that pushed this one up the list.
</example>
<task>Write an announcement for our new CSV export feature. It supports filtering by date range and custom column selection.</task>Claude will match the tone (casual, direct, specific), the length (short), and the format (no bullet points, no headers, conversational sign-off) from the example.
How many examples do you need?
For style matching: one or two is usually enough. For classification tasks: three to five examples that cover the range of categories. For complex structured outputs: at least one example of the full structure you want.
More examples help when the pattern is subtle or the categories are ambiguous. For most tasks, one good example is enough to get 80% of the way there.
How MindStudio Lets You Apply These Rules at Scale
Knowing these rules is useful. Applying them consistently across every Claude interaction — especially when you’re building tools for a team or automating workflows — is where it gets harder.
That’s where MindStudio fits. It’s a no-code platform for building AI agents, and it gives you a dedicated system prompt layer where you can encode all six of these rules once and have them apply to every request automatically. You can set roles, add negative prompting instructions, specify effort levels, and include few-shot examples — all in a persistent system prompt that runs before every user input.
Because MindStudio has 200+ AI models available out of the box, you can also run the same workflow across multiple Claude models (or switch between Claude and other models entirely) without touching any code. This makes it easy to test whether extended thinking is actually improving your results for a given task, or whether a simpler model configuration produces the same output at lower cost.
If you’re using Claude for anything beyond one-off queries — content generation, data extraction, internal tools, customer-facing apps — building it as a MindStudio agent means your prompting rules are baked in rather than manually applied each time.
You can try MindStudio free at mindstudio.ai.
Common Mistakes to Avoid
Beyond the six rules, a few patterns consistently produce poor results with Claude.
Over-prompting simple tasks. Adding role framing, XML tags, and effort levels to a prompt that just needs a one-line answer creates noise and slows things down. Match your prompt complexity to the task.
Contradictory instructions. Telling Claude to “be concise” but also to “cover all relevant considerations” will produce something in between — which satisfies neither instruction. Pick one direction.
Relying on system prompts to fix bad user prompts. System prompts set defaults and constraints. They can’t compensate for a user input that provides no context. If you’re building a tool for others, design the interface to capture necessary context before it reaches the model.
Iterating on output instead of the prompt. If Claude gives you something wrong three times, the issue is almost always the prompt, not random variation. Fix the prompt.
FAQ
What is Claude Fable 5?
Claude Fable 5 is a model in Anthropic’s Claude lineup. Like other Claude models, it responds to natural language instructions and can be prompted using Anthropic’s documented best practices — including role assignment, negative prompting, effort level specification, and few-shot examples.
What is the difference between Claude Sonnet, Opus, and extended thinking?
Claude models sit on a spectrum of capability and compute cost. Opus is Anthropic’s most capable model, designed for complex reasoning tasks. Extended thinking is a mode where Claude reasons through problems step by step before generating a response — it’s associated with Opus-tier processing. For most tasks, using a mid-tier model without extended thinking produces faster, cheaper results that are just as good.
Does prompt engineering still matter with newer Claude models?
Yes. Newer models are better at interpreting ambiguous prompts, but they still respond significantly better to well-structured, specific instructions. Prompting practices like role assignment, negative prompting, and few-shot examples produce measurable improvements in output quality regardless of model version. Anthropic’s own engineers still use these techniques internally.
How do I stop Claude from adding unnecessary caveats?
Add an explicit negative prompt instruction: “Do not add disclaimers, caveats, or recommendations to consult a professional.” Place this in your system prompt so it applies to every response. For one-off prompts, include it at the end of your instruction block.
When should I use XML tags in a Claude prompt?
Use XML tags any time your prompt contains more than one distinct component — instructions, content, examples, context. Tags remove ambiguity about what each part of the prompt is, which consistently improves output quality. For short, single-component prompts, tags are optional.
What is few-shot prompting and when does it help?
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
Few-shot prompting means providing one or more examples of the output you want before asking for the actual output. It’s most useful when you need Claude to match a specific style, format, or tone that’s hard to describe in words. One well-chosen example is usually enough for style matching; classification tasks benefit from three to five examples.
Key Takeaways
- Assign Claude a specific role before giving it a task — the role frames how it interprets and answers the request.
- Use negative prompting to remove default behaviors like caveats, preamble, and padding.
- Specify effort level explicitly so Claude knows whether to give you a sentence or an analysis.
- Avoid framing simple tasks as reasoning problems — it triggers extended thinking when you don’t need it.
- Use XML tags to separate instructions from content in complex prompts.
- Show Claude an example of what you want; descriptions are ambiguous, examples aren’t.
These techniques work across Claude models and compound well with each other. The most effective prompts tend to use several at once: a role, a negative instruction, an effort level, and an example. That combination consistently produces output that needs less iteration to get right.
If you’re building anything on top of Claude — from internal tools to customer-facing applications — MindStudio makes it easy to encode these rules into reusable agents without writing any infrastructure code. The no-code agent builder handles system prompts, model routing, and integrations so you can focus on what the agent actually does.
