Claude 4 prompting techniques for advanced reasoning models

Executive overview

Advanced reasoning models like Claude 4 follow instructions far more precisely than older models, making prompt wording more consequential than ever. Tactics that worked before — shouting in all caps, using negations, repeating instructions — are now counterproductive.

Five techniques extract the most from these models: specificity with modifiers, curated examples, XML delimiters, positive instruction framing, and leading by example in your own prompt style.

Write prompts the way you want the model to respond — structure, tone, and all.

Specificity and modifiers

• Add scope modifiers to expand output: "include as many relevant features and interactions as possible."
• Add depth modifiers to push beyond defaults: "go beyond the basics."
• Add completeness modifiers to signal thoroughness: "fully featured implementation."
• Words matter more now because the model executes instructions precisely rather than approximating intent.

Using examples effectively (few-shot prompting)

• Embed examples near the end of your prompt to anchor expected output format and behavior.
• Scrutinise every example — small nuances propagate into the model's output.
• Examples should encourage desired behavior and implicitly discourage what you want to avoid.

XML tags as delimiters

• Use XML tags to segment a prompt so the model knows which part is which (system context, user task, examples).
• Instruct the model to wrap its own output in tags — useful for parsing responses programmatically via API.
• Applies to all advanced reasoning models, not just Claude.

State what you want, not what to avoid

• Older models needed negation ("never use markdown"); current models respond better to positive instruction.
• Replace "do not use markdown" with "your response should be composed of smoothly flowing prose."
• Precise positive framing removes ambiguity and leverages the model's instruction-following strength.

Lead by example in your prompt

• Claude mirrors the style and structure of the prompt it receives.
• If you want prose output, write the prompt in prose — reduce markdown in the prompt itself.
• Style, tone, and formatting in your input influence the model's output, not just content instructions.

Improved visuals in UI generation

• Use encouragement phrases: "don't hold back, give it your all."
• Add UI-specific modifiers: hover states, transitions, micro-interactions, hierarchy, contrast, balance, movement.
• These nudges are more effective now because the model follows them literally.

Cleaning up agentic breadcrumbs

• Claude 4 in agentic/coding tasks creates temporary markdown files as scratch pads during execution.
• These accumulate and clutter the repository once the task is complete.
• Add a project rule: "if you create any temporary files or helper files for iteration, remove them at the end of the task."
• Works in Cursor, Windsurf, and other coding environments that support project rules.

Enforcing iterative thinking

• Claude 4 supports think-act-reflect loops: use a tool, reason over the result, pick the next tool.
• Bake this into system prompts: "after receiving tool results, carefully reflect on their quality and determine optimal next steps before proceeding."
• Effective for research and coding tasks requiring multi-step reasoning.
• Can be added to Claude Projects, GPT Projects, or Gemini Gems system instructions.

Parallel tool calling (superposition)

• Older models called tools sequentially — call, wait, call, wait — wasting context on idle time.
• Claude 4 can invoke multiple tools simultaneously and synthesise all responses at once.
• Prompt enforcement: "whenever you need to perform multiple independent operations, invoke all relevant tools simultaneously rather than sequentially."
• Improves complex task performance and efficiency.