Part 1: What Actually Works

1. Be Clear and Direct

This sounds obvious, but it's the single highest-impact prompting technique and the one most people skip. Anthropic's own documentation puts it simply: "Show your prompt to a colleague with minimal context. If they'd be confused, Claude will be too." Vague prompts produce vague outputs. Specific prompts produce specific outputs. The relationship is almost linear.

Instead of "help me with my code," say "refactor this Python function to use list comprehensions instead of for-loops, and add type hints to the parameters." Instead of "write something about climate change," say "write a 300-word summary of how carbon pricing mechanisms work, aimed at someone with no economics background." The more precisely you describe what you want, the less likely you are to need follow-up corrections.

This also applies to format. If you want a bullet list, say so. If you want a table, say so. If you want the answer in JSON, say so and provide the schema. Claude will match whatever structure you specify — but it can't read your mind about structure any more than a human colleague could.

2. Use Examples

Few-shot prompting — giving Claude 2–5 examples of the input-output pattern you want — is one of the most reliable ways to control output format, tone, and structure. It works because showing is more precise than telling. You can write a paragraph describing the format you want, or you can just show three examples and let Claude infer the pattern. The examples approach is faster to write and produces more consistent results.

This is especially useful for tasks with specific formatting requirements: structured data extraction, classification with a fixed label set, or writing in a particular style. If you want Claude to generate product descriptions in a specific brand voice, one example is worth a hundred words of style instructions. 

Here are examples of the format I need:

Input: "The quarterly earnings exceeded expectations by 12%"
Output: {"sentiment": "positive", "topic": "earnings", "confidence": 0.95}

Input: "Supply chain delays are expected to continue through Q3"
Output: {"sentiment": "negative", "topic": "supply_chain", "confidence": 0.88}

Now classify this:
Input: "The new product launch has been postponed indefinitely"

3. Structure with XML Tags

Claude is specifically trained to work well with XML-style tags for structuring prompts. This is one of the areas where Claude genuinely differs from other models. When you have a complex prompt with multiple types of information — instructions, context documents, examples, and a question — wrapping each in descriptive tags eliminates ambiguity about where one section ends and another begins. 

<instructions>
Summarize the following document in 3 bullet points.
Focus on actionable findings only.
</instructions>

<document>
[your document here]
</document>

<output_format>
Return exactly 3 bullet points, each starting with a verb.
</output_format>

This matters most when prompts get long. For a simple one-liner, tags are overkill. But once you're passing in documents, reference material, or multiple sets of instructions, XML tags help Claude parse the prompt unambiguously. They also make your prompts easier for humans to read and maintain, which is a practical benefit if you're iterating on prompts in a team.

4. Give Context for Why

Telling Claude what you want is necessary. Telling it why you want it often produces better results. "Summarize this document" and "summarize this document for a board meeting where the CEO needs to make a go/no-go decision on the project" will produce noticeably different outputs. The second version gives Claude enough context to prioritize the right information and frame it appropriately.

This works because context constrains the space of reasonable outputs. Without context, Claude has to guess at what matters. With context, it can make informed choices about emphasis, detail level, and tone. This is especially useful for writing tasks, analysis, and any situation where "correct" depends on who's reading the output.

5. Put Long Documents Before Your Question

When your prompt includes a long document (20k+ tokens), place the document before your instructions or questions. This is a practical concern about how attention works in transformer models: information at the very beginning and very end of the context tends to be weighted more heavily than information buried in the middle. Putting the document first and your question last means the question is fresh when Claude starts generating.

For extremely long contexts, consider telling Claude what to look for before the document as well: "The following is a 50-page contract. After reading it, identify all clauses related to termination rights." This primes the model to attend to the relevant sections during its forward pass, rather than processing the entire document with no focus.

6. Use the System Prompt for Role and Constraints

If you're using the API, the system prompt is the right place for persistent instructions: role definitions, output constraints, tone guidelines, and behavioral rules. Claude gives strong weight to system prompt instructions, more so in recent model versions. A role set in the system prompt focuses Claude's behavior more effectively than the same role set in the user message.

Keep the system prompt focused and concise. Anthropic's documentation notes that overly complex system prompts can cause Claude to overthink or overtrigger on certain behaviors. If you find the model being excessively cautious or verbose, the system prompt is usually the first place to look. Write it like you'd write a brief for a capable colleague: clear scope, clear constraints, and trust them to handle the details.

7. Let Claude Think

For complex reasoning tasks — math, logic, multi-step analysis, code debugging — giving Claude space to reason before answering improves accuracy. The newer Claude models (Opus 4.5/4.6, Sonnet 4.5/4.6) support extended thinking, where the model can do internal reasoning before producing a visible response. When available, this is the best approach for hard problems: you don't need to prompt for it manually, you just enable it in the API.

Even without extended thinking, a prompt like "reason through this carefully before answering" can help on complex tasks. But Anthropic's current guidance is interesting: "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." In other words, tell it to think hard, but don't micromanage the thinking process.

Part 2: What to Stop Doing

8. Stop Saying "Think Step by Step"

This was genuinely useful with GPT-3 and early GPT-4. It's been cargo-culted into nearly every prompting guide since. For current Claude models, it's largely unnecessary and sometimes counterproductive. Modern Claude models are already trained to reason through complex problems. Forcing a specific step-by-step format can actually constrain the model's reasoning into a structure that doesn't fit the problem.

If you want Claude to show its reasoning, say "show your reasoning" or "explain how you arrived at your answer." If you want it to reason more carefully, say "think thoroughly" or enable extended thinking. But the specific incantation "let's think step by step" is a relic from a different model generation. It won't break anything, but it's taking up tokens for minimal benefit.

9. Stop Threatening and Bribing the Model

"I'll tip you $100 for a great answer." "Your job depends on getting this right." "A person's life is at stake." These prompts circulate on social media with claims of improved output quality. They do not work in any mechanistically meaningful way. Claude does not have financial motivations, job security concerns, or emotional responses to threats. It processes tokens.

If these prompts occasionally seem to produce better results, it's because they add emphasis — the model may attend more to a task described as important. But you can achieve the same effect far more reliably by being specific about your requirements and quality criteria. "This needs to be production-ready code with error handling and edge case coverage" is actionable. "I'll fire you if this is wrong" is not.

10. Stop Adding "Take a Deep Breath"

This one originated from a 2023 paper that showed marginal improvements on math benchmarks with certain emotional prompts. It got wildly overgeneralized. Claude does not have a nervous system. It does not experience stress. It cannot calm down by breathing. The phrase is meaningless noise in the prompt.

The underlying intuition — that priming the model to be careful can help — is not entirely wrong. But "be thorough and precise" communicates the same intent in words the model can actually act on. If you're spending tokens on emotional regulation for a language model, those tokens would serve you better as an additional example or a clearer description of the output format.

11. Stop Repeating Instructions

Saying the same instruction three times does not make Claude follow it three times as hard. The model processes the full context and attends to instructions based on clarity and positioning, not repetition count. Repeating yourself wastes tokens, inflates costs, and makes your prompt harder to maintain.

If Claude isn't following an instruction, the fix is rarely to repeat it. More often, the instruction is ambiguous, contradicts something else in the prompt, or is buried in a wall of text. Move it to a more prominent position (beginning or end of the prompt), use XML tags to separate it, or rephrase it more specifically. One clear instruction outperforms three vague ones.

12. Stop Writing Novels in Your System Prompt

A 2,000-word system prompt full of edge cases, ALL-CAPS warnings, and nested conditionals is harder for Claude to follow than a focused 200-word one. Long system prompts also increase the chance of contradictory instructions, where one rule tells the model to be concise and another asks for detailed explanations. When instructions conflict, the model has to pick one, and you may not like which it chooses.

Anthropic specifically notes that Claude Opus 4.5 and 4.6 are more responsive to system prompts than previous models. Prompts that were designed to be aggressive to ensure compliance may now cause overtriggering. The fix is to dial back, not add more. Write the system prompt that a sharp colleague would need to do the job well, and resist the urge to add instructions for every hypothetical edge case.

13. Stop Using "You Are an Expert" as a Magic Spell

"You are a world-class expert in X" appears in nearly every prompting template online. It does have a mild effect — it sets a tone and frames the expected quality level. Anthropic confirms that setting a role in the system prompt focuses Claude's behavior. But it's a framing device, not a competence upgrade. Saying "you are a world-class Python developer" does not give Claude access to Python knowledge it wouldn't otherwise have.

Where role-setting actually helps is in tone and approach. "You are a senior engineer reviewing a pull request" produces different output than "you are a patient tutor explaining code to a beginner" — because the role implies a different communication style and level of detail. Use roles for framing, not as a substitute for clear instructions. And keep it to one sentence in the system prompt — you don't need a character backstory.

The Efficient Prompting Mindset

The most effective prompts share three properties: they're specific about what they want, they provide enough context for Claude to make good decisions, and they don't waste tokens on things that don't affect the output. Every token in your prompt costs money (if you're on the API) and attention (in the model's context window). Treat tokens like you'd treat words in a technical specification: every one should earn its place.

The common thread through the anti-patterns above is magical thinking — the belief that specific phrases have special power over the model. They don't. Claude processes natural language and follows instructions. The better your instructions, the better the output. There is no incantation that substitutes for clarity.

If you take one thing from this article: before adding anything to a prompt, ask whether it gives Claude information it can act on. Examples, context, format specifications, and constraints all qualify. Emotional appeals, threats, repetition, and breathing exercises do not. Spend your tokens on the things that work.

Resources